5. Advanced Recipes

The following recipes involve advanced techniques and may not be “beginner friendly”. Please feel free to suggest enhancements or additional recipes.

5.1. Unencoded image capture (YUV format)

If you want images captured without loss of detail (due to JPEG’s lossy compression), you are probably better off exploring PNG as an alternate image format (PNG uses lossless compression). However, some applications (particularly scientific ones) simply require the image data in numeric form. For this, the 'yuv' format is provided:

import time
import picamera

with picamera.PiCamera() as camera:
    camera.resolution = (100, 100)
    camera.start_preview()
    time.sleep(2)
    camera.capture('image.data', 'yuv')

The specific YUV format used is YUV420 (planar). This means that the Y (luminance) values occur first in the resulting data and have full resolution (one 1-byte Y value for each pixel in the image). The Y values are followed by the U (chrominance) values, and finally the V (chrominance) values. The UV values have one quarter the resolution of the Y components (4 1-byte Y values in a square for each 1-byte U and 1-byte V value). This is illustrated in the diagram below:

It is also important to note that when outputting to unencoded formats, the camera rounds the requested resolution. The horizontal resolution is rounded up to the nearest multiple of 32 pixels, while the vertical resolution is rounded up to the nearest multiple of 16 pixels. For example, if the requested resolution is 100x100, the capture will actually contain 128x112 pixels worth of data, but pixels beyond 100x100 will be uninitialized.

Given that the YUV420 format contains 1.5 bytes worth of data for each pixel (a 1-byte Y value for each pixel, and 1-byte U and V values for every 4 pixels), and taking into account the resolution rounding, the size of a 100x100 YUV capture will be:

The first 14336 bytes of the data (128*112) will be Y values, the next 3584 bytes (128*112/4) will be U values, and the final 3584 bytes will be the V values.

The following code demonstrates capturing YUV image data, loading the data into a set of numpy arrays, and converting the data to RGB format in an efficient manner:

from __future__ import division

import time
import picamera
import numpy as np

width = 100
height = 100
stream = open('image.data', 'w+b')
# Capture the image in YUV format
with picamera.PiCamera() as camera:
    camera.resolution = (width, height)
    camera.start_preview()
    time.sleep(2)
    camera.capture(stream, 'yuv')
# Rewind the stream for reading
stream.seek(0)
# Calculate the actual image size in the stream (accounting for rounding
# of the resolution)
fwidth = (width + 31) // 32 * 32
fheight = (height + 15) // 16 * 16
# Load the Y (luminance) data from the stream
Y = np.fromfile(stream, dtype=np.uint8, count=fwidth*fheight).\
        reshape((fheight, fwidth))
# Load the UV (chrominance) data from the stream, and double its size
U = np.fromfile(stream, dtype=np.uint8, count=(fwidth//2)*(fheight//2)).\
        reshape((fheight//2, fwidth//2)).\
        repeat(2, axis=0).repeat(2, axis=1)
V = np.fromfile(stream, dtype=np.uint8, count=(fwidth//2)*(fheight//2)).\
        reshape((fheight//2, fwidth//2)).\
        repeat(2, axis=0).repeat(2, axis=1)
# Stack the YUV channels together, crop the actual resolution, convert to
# floating point for later calculations, and apply the standard biases
YUV = np.dstack((Y, U, V))[:height, :width, :].astype(np.float)
YUV[:, :, 0]  = YUV[:, :, 0]  - 16   # Offset Y by 16
YUV[:, :, 1:] = YUV[:, :, 1:] - 128  # Offset UV by 128
# YUV conversion matrix from ITU-R BT.601 version (SDTV)
#              Y       U       V
M = np.array([[1.164,  0.000,  1.596],    # R
              [1.164, -0.392, -0.813],    # G
              [1.164,  2.017,  0.000]])   # B
# Take the dot product with the matrix to produce RGB output, clamp the
# results to byte range and convert to bytes
RGB = YUV.dot(M.T).clip(0, 255).astype(np.uint8)

Note

You may note that we are using open() in the code above instead of io.open() as in the other examples. This is because numpy’s numpy.fromfile() method annoyingly only accepts “real” file objects.

This recipe is now encapsulated in the PiYUVArray class in the picamera.array module, which means the same can be achieved as follows:

import time
import picamera
import picamera.array

with picamera.PiCamera() as camera:
    with picamera.array.PiYUVArray(camera) as stream:
        camera.resolution = (100, 100)
        camera.start_preview()
        time.sleep(2)
        camera.capture(stream, 'yuv')
        # Show size of YUV data
        print(stream.array.shape)
        # Show size of RGB converted data
        print(stream.rgb_array.shape)

Alternatively, see Unencoded image capture (RGB format) for a method of having the camera output RGB data directly.

Note

Capturing so-called “raw” formats ('yuv', 'rgb', etc.) does not provide the raw bayer data from the camera’s sensor. Rather, it provides access to the image data after GPU processing, but before format encoding (JPEG, PNG, etc). Currently, the only method of accessing the raw bayer data is via the bayer parameter to the capture() method. See Raw Bayer data captures for more information.

Changed in version 1.0: The raw_format attribute is now deprecated, as is the 'raw' format specification for the capture() method. Simply use the 'yuv' format instead, as shown in the code above.

Changed in version 1.5: Added note about new picamera.array module.

5.2. Unencoded image capture (RGB format)

The RGB format is rather larger than the YUV format discussed in the section above, but is more useful for most analyses. To have the camera produce output in RGB format, you simply need to specify 'rgb' as the format for the capture() method instead:

import time
import picamera

with picamera.PiCamera() as camera:
    camera.resolution = (100, 100)
    camera.start_preview()
    time.sleep(2)
    camera.capture('image.data', 'rgb')

The size of RGB data can be calculated similarly to YUV captures. Firstly round the resolution appropriately (see Unencoded image capture (YUV format) for the specifics), then multiply the number of pixels by 3 (1 byte of red, 1 byte of green, and 1 byte of blue intensity). Hence, for a 100x100 capture, the amount of data produced is:

The resulting RGB data is interleaved. That is to say that the red, green and blue values for a given pixel are grouped together, in that order. The first byte of the data is the red value for the pixel at (0, 0), the second byte is the green value for the same pixel, and the third byte is the blue value for that pixel. The fourth byte is the red value for the pixel at (1, 0), and so on.

Loading the resulting RGB data into a numpy array is simple:

from __future__ import division

width = 100
height = 100
stream = open('image.data', 'w+b')
# Capture the image in RGB format
with picamera.PiCamera() as camera:
    camera.resolution = (width, height)
    camera.start_preview()
    time.sleep(2)
    camera.capture(stream, 'rgb')
# Rewind the stream for reading
stream.seek(0)
# Calculate the actual image size in the stream (accounting for rounding
# of the resolution)
fwidth = (width + 31) // 32 * 32
fheight = (height + 15) // 16 * 16
# Load the data in a three-dimensional array and crop it to the requested
# resolution
image = np.fromfile(stream, dtype=np.uint8).\
        reshape((fheight, fwidth, 3))[:height, :width, :]
# If you wish, the following code will convert the image's bytes into
# floating point values in the range 0 to 1 (a typical format for some
# sorts of analysis)
image = image.astype(np.float, copy=False)
image = image / 255.0

This recipe is now encapsulated in the PiRGBArray class in the picamera.array module, which means the same can be achieved as follows:

import time
import picamera
import picamera.array

with picamera.PiCamera() as camera:
    with picamera.array.PiRGBArray(camera) as stream:
        camera.resolution = (100, 100)
        camera.start_preview()
        time.sleep(2)
        camera.capture(stream, 'rgb')
        # Show size of RGB data
        print(stream.array.shape)

Note

RGB captures from the still port do not work at the full resolution of the camera (they result in an out of memory error). Either use YUV captures, or capture from the video port if you require full resolution.

Changed in version 1.0: The raw_format attribute is now deprecated, as is the 'raw' format specification for the capture() method. Simply use the 'rgb' format instead, as shown in the code above.

Changed in version 1.5: Added note about new picamera.array module.

5.3. Rapid capture and processing

The camera is capable of capturing a sequence of images extremely rapidly by utilizing its video-capture capabilities with a JPEG encoder (via the use_video_port parameter). However, there are several things to note about using this technique:

  • When using video-port based capture only the video recording area is captured; in some cases this may be smaller than the normal image capture area (see dicussion in Camera Modes).
  • No Exif information is embedded in JPEG images captured through the video-port.
  • Captures typically appear “granier” with this technique. Captures from the still port use a slower, more intensive denoise algorithm.

All capture methods support the use_video_port option, but the methods differ in their ability to rapidly capture sequential frames. So, whilst capture() and capture_continuous() both support use_video_port, capture_sequence() is by far the fastest method (because it does not re-initialize an encoder prior to each capture). Using this method, the author has managed 30fps JPEG captures at a resolution of 1024x768.

By default, capture_sequence() is particularly suited to capturing a fixed number of frames rapidly, as in the following example which captures a “burst” of 5 images:

import time
import picamera

with picamera.PiCamera() as camera:
    camera.resolution = (1024, 768)
    camera.framerate = 30
    camera.start_preview()
    time.sleep(2)
    camera.capture_sequence([
        'image1.jpg',
        'image2.jpg',
        'image3.jpg',
        'image4.jpg',
        'image5.jpg',
        ])

We can refine this slightly by using a generator expression to provide the filenames for processing instead of specifying every single filename manually:

import time
import picamera

frames = 60

with picamera.PiCamera() as camera:
    camera.resolution = (1024, 768)
    camera.framerate = 30
    camera.start_preview()
    # Give the camera some warm-up time
    time.sleep(2)
    start = time.time()
    camera.capture_sequence([
        'image%02d.jpg' % i
        for i in range(frames)
        ], use_video_port=True)
    finish = time.time()
print('Captured %d frames at %.2ffps' % (
    frames,
    frames / (finish - start)))

However, this still doesn’t let us capture an arbitrary number of frames until some condition is satisfied. To do this we need to use a generator function to provide the list of filenames (or more usefully, streams) to the capture_sequence() method:

import time
import picamera

frames = 60

def filenames():
    frame = 0
    while frame < frames:
        yield 'image%02d.jpg' % frame
        frame += 1

with picamera.PiCamera() as camera:
    camera.resolution = (1024, 768)
    camera.framerate = 30
    camera.start_preview()
    # Give the camera some warm-up time
    time.sleep(2)
    start = time.time()
    camera.capture_sequence(filenames(), use_video_port=True)
    finish = time.time()
print('Captured %d frames at %.2ffps' % (
    frames,
    frames / (finish - start)))

The major issue with capturing this rapidly is firstly that the Raspberry Pi’s IO bandwidth is extremely limited and secondly that, as a format, JPEG is considerably less efficient than the H.264 video format (which is to say that, for the same number of bytes, H.264 will provide considerably better quality over the same number of frames). At higher resolutions (beyond 800x600) you are likely to find you cannot sustain 30fps captures to the Pi’s SD card for very long (before exhausting the disk cache).

If you are intending to perform processing on the frames after capture, you may be better off just capturing video and decoding frames from the resulting file rather than dealing with individual JPEG captures. Alternatively, you may wish to investigate sending the data over the network (which typically has more bandwidth available than the SD card interface) and having another machine perform any required processing. However, if you can perform your processing fast enough, you may not need to involve the disk or network at all. Using a generator function, we can maintain a queue of objects to store the captures, and have parallel threads accept and process the streams as captures come in. Provided the processing runs at a faster frame rate than the captures, the encoder won’t stall:

import io
import time
import threading
import picamera

# Create a pool of image processors
done = False
lock = threading.Lock()
pool = []

class ImageProcessor(threading.Thread):
    def __init__(self):
        super(ImageProcessor, self).__init__()
        self.stream = io.BytesIO()
        self.event = threading.Event()
        self.terminated = False
        self.start()

    def run(self):
        # This method runs in a separate thread
        global done
        while not self.terminated:
            # Wait for an image to be written to the stream
            if self.event.wait(1):
                try:
                    self.stream.seek(0)
                    # Read the image and do some processing on it
                    #Image.open(self.stream)
                    #...
                    #...
                    # Set done to True if you want the script to terminate
                    # at some point
                    #done=True
                finally:
                    # Reset the stream and event
                    self.stream.seek(0)
                    self.stream.truncate()
                    self.event.clear()
                    # Return ourselves to the pool
                    with lock:
                        pool.append(self)

def streams():
    while not done:
        with lock:
            if pool:
                processor = pool.pop()
            else:
                processor = None
        if processor:
            yield processor.stream
            processor.event.set()
        else:
            # When the pool is starved, wait a while for it to refill
            time.sleep(0.1)

with picamera.PiCamera() as camera:
    pool = [ImageProcessor() for i in range(4)]
    camera.resolution = (640, 480)
    camera.framerate = 30
    camera.start_preview()
    time.sleep(2)
    camera.capture_sequence(streams(), use_video_port=True)

# Shut down the processors in an orderly fashion
while pool:
    with lock:
        processor = pool.pop()
    processor.terminated = True
    processor.join()

5.4. Rapid capture and streaming

Following on from Rapid capture and processing, we can combine the video-port capture technique with Capturing to a network stream. The server side script doesn’t change (it doesn’t really care what capture technique is being used - it just reads JPEGs off the wire). The changes to the client side script can be minimal at first - just add use_video_port=True to the capture_continuous() call:

import io
import socket
import struct
import time
import picamera

client_socket = socket.socket()
client_socket.connect(('my_server', 8000))
connection = client_socket.makefile('wb')
try:
    with picamera.PiCamera() as camera:
        camera.resolution = (640, 480)
        camera.framerate = 30
        time.sleep(2)
        start = time.time()
        stream = io.BytesIO()
        # Use the video-port for captures...
        for foo in camera.capture_continuous(stream, 'jpeg',
                                             use_video_port=True):
            connection.write(struct.pack('<L', stream.tell()))
            connection.flush()
            stream.seek(0)
            connection.write(stream.read())
            if time.time() - start > 30:
                break
            stream.seek(0)
            stream.truncate()
    connection.write(struct.pack('<L', 0))
finally:
    connection.close()
    client_socket.close()

Using this technique, the author can manage about 10fps of streaming at 640x480 on firmware #685. One deficiency of the script above is that it interleaves capturing images with sending them over the wire (although we deliberately don’t flush on sending the image data). Potentially, it would be more efficient to permit image capture to occur simultaneously with image transmission. We can attempt to do this by utilizing the background threading techniques from the final example in Rapid capture and processing:

import io
import socket
import struct
import time
import threading
import picamera

client_socket = socket.socket()
client_socket.connect(('spider', 8000))
connection = client_socket.makefile('wb')
try:
    connection_lock = threading.Lock()
    pool_lock = threading.Lock()
    pool = []

    class ImageStreamer(threading.Thread):
        def __init__(self):
            super(ImageStreamer, self).__init__()
            self.stream = io.BytesIO()
            self.event = threading.Event()
            self.terminated = False
            self.start()

        def run(self):
            # This method runs in a background thread
            while not self.terminated:
                # Wait for the image to be written to the stream
                if self.event.wait(1):
                    try:
                        with connection_lock:
                            connection.write(struct.pack('<L', self.stream.tell()))
                            connection.flush()
                            self.stream.seek(0)
                            connection.write(self.stream.read())
                    finally:
                        self.stream.seek(0)
                        self.stream.truncate()
                        self.event.clear()
                        with pool_lock:
                            pool.append(self)

    count = 0
    start = time.time()
    finish = time.time()

    def streams():
        global count, finish
        while finish - start < 30:
            with pool_lock:
                if pool:
                    streamer = pool.pop()
                else:
                    streamer = None
            if streamer:
                yield streamer.stream
                streamer.event.set()
                count += 1
            else:
                # When the pool is starved, wait a while for it to refill
                time.sleep(0.1)
            finish = time.time()

    with picamera.PiCamera() as camera:
        pool = [ImageStreamer() for i in range(4)]
        camera.resolution = (640, 480)
        camera.framerate = 30
        time.sleep(2)
        start = time.time()
        camera.capture_sequence(streams(), 'jpeg', use_video_port=True)

    # Shut down the streamers in an orderly fashion
    while pool:
        streamer = pool.pop()
        streamer.terminated = True
        streamer.join()

    # Write the terminating 0-length to the connection to let the server
    # know we're done
    with connection_lock:
        connection.write(struct.pack('<L', 0))

finally:
    connection.close()
    client_socket.close()

print('Sent %d images in %d seconds at %.2ffps' % (
    count, finish-start, count / (finish-start)))

On the same firmware, the above script achieves about 15fps. It is possible the new high framerate modes may achieve more (the fact that 15fps is half of the specified 30fps framerate suggests some stall on every other frame).

5.5. Capturing images whilst recording

The camera is capable of capturing still images while it is recording video. However, if one attempts this using the stills capture mode, the resulting video will have dropped frames during the still image capture. This is because images captured via the still port require a mode change, causing the dropped frames (this is the flicker to a higher resolution that one sees when capturing while a preview is running).

However, if the use_video_port parameter is used to force a video-port based image capture (see Rapid capture and processing) then the mode change does not occur, and the resulting video should not have dropped frames, assuming the image can be produced before the next video frame is due:

import picamera

with picamera.PiCamera() as camera:
    camera.resolution = (800, 600)
    camera.start_preview()
    camera.start_recording('foo.h264')
    camera.wait_recording(10)
    camera.capture('foo.jpg', use_video_port=True)
    camera.wait_recording(10)
    camera.stop_recording()

The above code should produce a 20 second video with no dropped frames, and a still frame from 10 seconds into the video. Higher resolutions or non-JPEG image formats may still cause dropped frames (only JPEG encoding is hardware accelerated).

5.6. Recording at multiple resolutions

The camera is capable of recording multiple streams at different resolutions simultaneously by use of the video splitter. This is probably most useful for performing analysis on a low-resolution stream, while simultaneously recording a high resolution stream for storage or viewing.

The following simple recipe demonstrates using the splitter_port parameter of the start_recording() method to begin two simultaneous recordings, each with a different resolution:

import picamera

with picamera.PiCamera() as camera:
    camera.resolution = (1024, 768)
    camera.framerate = 30
    camera.start_recording('highres.h264')
    camera.start_recording('lowres.h264', splitter_port=2, resize=(320, 240))
    camera.wait_recording(30)
    camera.stop_recording(splitter_port=2)
    camera.stop_recording()

There are 4 splitter ports in total that can be used (numbered 0, 1, 2, and 3). The video recording methods default to using splitter port 1, while the image capture methods default to splitter port 0 (when the use_video_port parameter is also True). A splitter port cannot be simultaneously used for video recording and image capture so you are advised to avoid splitter port 0 for video recordings unless you never intend to capture images whilst recording.

New in version 1.3.

5.7. Recording motion vector data

The Pi’s camera is capable of outputting the motion vector estimates that the camera’s H.264 encoder calculates while generating compressed video. These can be directed to a separate output file (or file-like object) with the motion_output parameter of the start_recording() method. Like the normal output parameter this accepts a string representing a filename, or a file-like object:

import picamera

with picamera.PiCamera() as camera:
    camera.resolution = (640, 480)
    camera.framerate = 30
    camera.start_recording('motion.h264', motion_output='motion.data')
    camera.wait_recording(10)
    camera.stop_recording()

Motion data is calculated at the macro-block level (an MPEG macro-block represents a 16x16 pixel region of the frame), and includes one extra column of data. Hence, if the camera’s resolution is 640x480 (as in the example above) there will be 41 columns of motion data ((640 / 16) + 1), in 30 rows (480 / 16).

Motion data values are 4-bytes long, consisting of a signed 1-byte x vector, a signed 1-byte y vector, and an unsigned 2-byte SAD (Sum of Absolute Differences) value for each macro-block. Hence in the example above, each frame will generate 4920 bytes of motion data (41 * 30 * 4). Assuming the data contains 300 frames (in practice it may contain a few more) the motion data should be 1,476,000 bytes in total.

The following code demonstrates loading the motion data into a three-dimensional numpy array. The first dimension represents the frame, with the latter two representing rows and finally columns. A structured data-type is used for the array permitting easy access to x, y, and SAD values:

from __future__ import division

import numpy as np

width = 640
height = 480
cols = (width + 15) // 16
cols += 1 # there's always an extra column
rows = (height + 15) // 16

motion_data = np.fromfile(
    'motion.data', dtype=[
        ('x', 'i1'),
        ('y', 'i1'),
        ('sad', 'u2'),
        ])
frames = motion_data.shape[0] // (cols * rows * motion_data.dtype.itemsize)
motion_data = motion_data.reshape((frames, rows, cols))

# Access the data for the first frame
motion_data[0]

# Access just the x-vectors from the fifth frame
motion_data[4]['x']

# Access SAD values for the tenth frame
motion_data[9]['sad']

You can calculate the amount of motion the vector represents simply by calculating the magnitude of the vector with Pythagoras’ theorem. The SAD (Sum of Absolute Differences) value can be used to determine how well the encoder thinks the vector represents the original reference frame.

The following code extends the example above to use PIL to produce a PNG image from the magnitude of each frame’s motion vectors:

from __future__ import division

import numpy as np
from PIL import Image

width = 640
height = 480
cols = (width + 15) // 16
rows = (height + 15) // 16

m = np.fromfile(
    'motion.data', dtype=[
        ('x', 'i1'),
        ('y', 'i1'),
        ('sad', 'u2'),
        ])
frames = m.shape[0] // (cols * rows * m.dtype.itemsize)
m = m.reshape((frames, rows, cols))

for frame in range(frames):
    data = np.sqrt(
        np.square(m[frame]['x'].astype(np.float)) +
        np.square(m[frame]['y'].astype(np.float))
        ).clip(0, 255).astype(np.uint8)
    img = Image.fromarray(data)
    filename = 'frame%03d.png' % frame
    print('Writing %s' % filename)
    img.save(filename)

You may wish to investigate the PiMotionArray class in the picamera.array module which simplifies the above recipes to the following:

import numpy as np
import picamera
import picamera.array
from PIL import Image

with picamera.PiCamera() as camera:
    with picamera.array.PiMotionArray(camera) as stream:
        camera.resolution = (640, 480)
        camera.framerate = 30
        camera.start_recording('/dev/null', format='h264', motion_output=stream)
        camera.wait_recording(10)
        camera.stop_recording()
        for frame in range(stream.array.shape[0]):
            data = np.sqrt(
                np.square(stream.array[frame]['x'].astype(np.float)) +
                np.square(stream.array[frame]['y'].astype(np.float))
                ).clip(0, 255).astype(np.uint8)
            img = Image.fromarray(data)
            filename = 'frame%03d.png' % frame
            print('Writing %s' % filename)
            img.save(filename)

Finally, the following command line can be used to generate an animation from the generated PNGs with ffmpeg (this will take a very long time on the Pi so you may wish to transfer the images to a faster machine for this step):

avconv -r 30 -i frame%03d.png -filter:v scale=640:480 -c:v libx264 motion.mp4

New in version 1.5.

5.8. Splitting to/from a circular stream

This example builds on the one in Recording to a circular stream and the one in Capturing images whilst recording to demonstrate the beginnings of a security application. As before, a PiCameraCircularIO instance is used to keep the last few seconds of video recorded in memory. While the video is being recorded, video-port-based still captures are taken to provide a motion detection routine with some input (the actual motion detection algorithm is left as an exercise for the reader).

Once motion is detected, the last 10 seconds of video are written to disk, and video recording is split to another disk file to proceed until motion is no longer detected. Once motion is no longer detected, we split the recording back to the in-memory ring-buffer:

import io
import random
import picamera
from PIL import Image

prior_image = None

def detect_motion(camera):
    global prior_image
    stream = io.BytesIO()
    camera.capture(stream, format='jpeg', use_video_port=True)
    stream.seek(0)
    if prior_image is None:
        prior_image = Image.open(stream)
        return False
    else:
        current_image = Image.open(stream)
        # Compare current_image to prior_image to detect motion. This is
        # left as an exercise for the reader!
        result = random.randint(0, 10) == 0
        # Once motion detection is done, make the prior image the current
        prior_image = current_image
        return result

def write_video(stream):
    # Write the entire content of the circular buffer to disk. No need to
    # lock the stream here as we're definitely not writing to it
    # simultaneously
    with io.open('before.h264', 'wb') as output:
        for frame in stream.frames:
            if frame.frame_type == picamera.PiVideoFrameType.sps_header:
                stream.seek(frame.position)
                break
        while True:
            buf = stream.read1()
            if not buf:
                break
            output.write(buf)
    # Wipe the circular stream once we're done
    stream.seek(0)
    stream.truncate()

with picamera.PiCamera() as camera:
    camera.resolution = (1280, 720)
    stream = picamera.PiCameraCircularIO(camera, seconds=10)
    camera.start_recording(stream, format='h264')
    try:
        while True:
            camera.wait_recording(1)
            if detect_motion(camera):
                print('Motion detected!')
                # As soon as we detect motion, split the recording to
                # record the frames "after" motion
                camera.split_recording('after.h264')
                # Write the 10 seconds "before" motion to disk as well
                write_video(stream)
                # Wait until motion is no longer detected, then split
                # recording back to the in-memory circular buffer
                while detect_motion(camera):
                    camera.wait_recording(1)
                print('Motion stopped!')
                camera.split_recording(stream)
    finally:
        camera.stop_recording()

This example also demonstrates writing the circular buffer to disk in an efficient manner using the read1() method (as opposed to read()).

Note

Note that read1() does not guarantee to return the number of bytes requested, even if they are available in the underlying stream; it simply returns as many as are available from a single chunk up to the limit specified.

New in version 1.0.

5.9. Custom outputs

All methods in the picamera library which accept a filename also accept file-like objects. Typically, this is only used with actual file objects, or with memory streams (like io.BytesIO). However, building a custom output object is extremely easy and in certain cases very useful. A file-like object (as far as picamera is concerned) is simply an object with a write method which must accept a single parameter consisting of a byte-string, and which can optionally return the number of bytes written. The object can optionally implement a flush method (which has no parameters), which will be called at the end of output.

Custom outputs are particularly useful with video recording as the custom output’s write method will be called (at least) once for every frame that is output, allowing you to implement code that reacts to each and every frame without going to the bother of a full custom encoder. However, one should bear in mind that because the write method is called so frequently, its implementation must be sufficiently rapid that it doesn’t stall the encoder (it must perform its processing and return before the next write is due to arrive).

The following trivial example demonstrates an incredibly simple custom output which simply throws away the output while counting the number of bytes that would have been written and prints this at the end of the output:

from __future__ import print_function

import picamera

class MyOutput(object):
    def __init__(self):
        self.size = 0

    def write(self, s):
        self.size += len(s)

    def flush(self):
        print('%d bytes would have been written' % self.size)

with picamera.PiCamera() as camera:
    camera.resolution = (640, 480)
    camera.framerate = 60
    camera.start_recording(MyOutput(), format='h264')
    camera.wait_recording(10)
    camera.stop_recording()

The following example shows how to use a custom output to construct a crude motion detection system. We construct a custom output object which is used as the destination for motion vector data (this is particularly simple as motion vector data always arrives as single chunks; frame data by contrast sometimes arrives in several separate chunks). The output object doesn’t actually write the motion data anywhere; instead it loads it into a numpy array and analyses whether there are any significantly large vectors in the data, printing a message to the console if there are. As we are not concerned with keeping the actual video output in this example, we use /dev/null as the destination for the video data:

from __future__ import division

import picamera
import numpy as np

motion_dtype = np.dtype([
    ('x', 'i1'),
    ('y', 'i1'),
    ('sad', 'u2'),
    ])

class MyMotionDetector(object):
    def __init__(self, camera):
        width, height = camera.resolution
        self.cols = (width + 15) // 16
        self.cols += 1 # there's always an extra column
        self.rows = (height + 15) // 16

    def write(self, s):
        # Load the motion data from the string to a numpy array
        data = np.fromstring(s, dtype=motion_dtype)
        # Re-shape it and calculate the magnitude of each vector
        data = data.reshape((self.rows, self.cols))
        data = np.sqrt(
            np.square(data['x'].astype(np.float)) +
            np.square(data['y'].astype(np.float))
            ).clip(0, 255).astype(np.uint8)
        # If there're more than 10 vectors with a magnitude greater
        # than 60, then say we've detected motion
        if (data > 60).sum() > 10:
            print('Motion detected!')
        # Pretend we wrote all the bytes of s
        return len(s)

with picamera.PiCamera() as camera:
    camera.resolution = (640, 480)
    camera.framerate = 30
    camera.start_recording(
        # Throw away the video data, but make sure we're using H.264
        '/dev/null', format='h264',
        # Record motion data to our custom output object
        motion_output=MyMotionDetector(camera)
        )
    camera.wait_recording(30)
    camera.stop_recording()

You may wish to investigate the classes in the picamera.array module which implement several custom outputs for analysis of data with numpy. In particular, the PiMotionAnalysis class can be used to remove much of the boiler plate code from the recipe above:

import picamera
import picamera.array
import numpy as np

class MyMotionDetector(picamera.array.PiMotionAnalysis):
    def analyse(self, a):
        a = np.sqrt(
            np.square(a['x'].astype(np.float)) +
            np.square(a['y'].astype(np.float))
            ).clip(0, 255).astype(np.uint8)
        # If there're more than 10 vectors with a magnitude greater
        # than 60, then say we've detected motion
        if (a > 60).sum() > 10:
            print('Motion detected!')

with picamera.PiCamera() as camera:
    camera.resolution = (640, 480)
    camera.framerate = 30
    camera.start_recording(
        '/dev/null', format='h264',
        motion_output=MyMotionDetector(camera)
        )
    camera.wait_recording(30)
    camera.stop_recording()

New in version 1.5.

5.10. Custom encoders

You can override and/or extend the encoder classes used during image or video capture. This is particularly useful with video capture as it allows you to run your own code in response to every frame, although naturally whatever code runs within the encoder’s callback has to be reasonably quick to avoid stalling the encoder pipeline.

Writing a custom encoder is quite a bit harder than writing a custom output and in most cases there’s little benefit. The only thing a custom encoder gives you that a custom output doesn’t is access to the buffer header flags. For many output formats (MJPEG and YUV for example), these won’t tell you anything interesting (i.e. they’ll simply indicate that the buffer contains a full frame and nothing else). Currently, the only format where the buffer header flags contain useful information is H.264. Even then, most of the information (I-frame, P-frame, motion information, etc.) would be accessible from the frame attribute which you could access from your custom output’s write method.

The encoder classes defined by picamera form the following hierarchy (shaded classes are actually instantiated by the implementation in picamera, white classes implement base functionality but aren’t technically “abstract”):

The following table details which PiCamera methods use which encoder classes, and which method they call to construct these encoders:

Method(s) Call Returns
capture() capture_continuous() capture_sequence() _get_image_encoder() PiCookedOneImageEncoder PiRawOneImageEncoder
capture_sequence() _get_images_encoder() PiCookedMultiImageEncoder PiRawMultiImageEncoder
start_recording() record_sequence() _get_video_encoder() PiCookedVideoEncoder PiRawVideoEncoder

It is recommended, particularly in the case of the image encoder classes, that you familiarize yourself with the specific function of these classes so that you can determine the best class to extend for your particular needs. You may find that one of the intermediate classes is a better basis for your own modifications.

In the following example recipe we will extend the PiCookedVideoEncoder class to store how many I-frames and P-frames are captured (the camera’s encoder doesn’t use B-frames):

import picamera
import picamera.mmal as mmal


# Override PiVideoEncoder to keep track of the number of each type of frame
class MyEncoder(picamera.PiCookedVideoEncoder):
    def start(self, output, motion_output=None):
        self.parent.i_frames = 0
        self.parent.p_frames = 0
        super(MyEncoder, self).start(output, motion_output)

    def _callback_write(self, buf):
        # Only count when buffer indicates it's the end of a frame, and
        # it's not an SPS/PPS header (..._CONFIG)
        if (
                (buf[0].flags & mmal.MMAL_BUFFER_HEADER_FLAG_FRAME_END) and
                not (buf[0].flags & mmal.MMAL_BUFFER_HEADER_FLAG_CONFIG)
            ):
            if buf[0].flags & mmal.MMAL_BUFFER_HEADER_FLAG_KEYFRAME:
                self.parent.i_frames += 1
            else:
                self.parent.p_frames += 1
        # Remember to return the result of the parent method!
        return super(MyEncoder, self)._callback_write(buf)


# Override PiCamera to use our custom encoder for video recording
class MyCamera(picamera.PiCamera):
    def __init__(self):
        super(MyCamera, self).__init__()
        self.i_frames = 0
        self.p_frames = 0

    def _get_video_encoder(
            self, camera_port, output_port, format, resize, **options):
        return MyEncoder(
                self, camera_port, output_port, format, resize, **options)


with MyCamera() as camera:
    camera.start_recording('foo.h264')
    camera.wait_recording(10)
    camera.stop_recording()
    print('Recording contains %d I-frames and %d P-frames' % (
            camera.i_frames, camera.p_frames))

Please note that the above recipe is flawed: PiCamera is capable of initiating multiple simultaneous recordings. If this were used with the above recipe, then each encoder would wind up incrementing the i_frames and p_frames attributes on the MyCamera instance leading to incorrect results.

New in version 1.5.

5.11. Raw Bayer data captures

The bayer parameter of the capture() method causes the raw Bayer data recorded by the camera’s sensor to be output as part of the image metadata.

Note

The bayer parameter only operates with the JPEG format, and only for captures from the still port (i.e. when use_video_port is False, as it is by default).

Raw Bayer data differs considerably from simple unencoded captures; it is the data recorded by the camera’s sensor prior to any GPU processing including auto white balance, vignette compensation, smoothing, down-scaling, etc. This also means:

  • Bayer data is always full resolution, regardless of the camera’s output resolution and any resize parameter.
  • Bayer data occupies the last 6,404,096 bytes of the output file. The first 32,768 bytes of this is header data which starts with the string 'BRCM'.
  • Bayer data consists of 10-bit values, because this is the sensitivity of the OV5647 sensor used by the Pi’s camera. The 10-bit values are organized as 4 8-bit values, followed by the low-order 2-bits of the 4 values packed into a fifth byte.
  • Bayer data is organized in a BGGR pattern (a minor variation of the common Bayer CFA). The raw data therefore has twice as many green pixels as red or blue and if viewed “raw” will look distinctly strange (too dark, too green, and with zippering effects along any straight edges).
  • To make a “normal” looking image from raw Bayer data you will need to perform de-mosaicing at the very least, and probably some form of color balance.

This (heavily commented) example script causes the camera to capture an image including the raw Bayer data. It then proceeds to unpack the Bayer data into a 3-dimensional numpy array representing the raw RGB data and finally performs a rudimentary de-mosaic step with weighted averages. A couple of numpy tricks are used to improve performance but bear in mind that all processing is happening on the CPU and will be considerably slower than normal image captures:

from __future__ import (
    unicode_literals,
    absolute_import,
    print_function,
    division,
    )


import io
import time
import picamera
import numpy as np
from numpy.lib.stride_tricks import as_strided

stream = io.BytesIO()
with picamera.PiCamera() as camera:
    # Let the camera warm up for a couple of seconds
    time.sleep(2)
    # Capture the image, including the Bayer data
    camera.capture(stream, format='jpeg', bayer=True)

# Extract the raw Bayer data from the end of the stream, check the
# header and strip if off before converting the data into a numpy array

data = stream.getvalue()[-6404096:]
assert data[:4] == 'BRCM'
data = data[32768:]
data = np.fromstring(data, dtype=np.uint8)

# The data consists of 1952 rows of 3264 bytes of data. The last 8 rows
# of data are unused (they only exist because the actual resolution of
# 1944 rows is rounded up to the nearest 16). Likewise, the last 24
# bytes of each row are unused (why?). Here we reshape the data and
# strip off the unused bytes

data = data.reshape((1952, 3264))[:1944, :3240]

# Horizontally, each row consists of 2592 10-bit values. Every four
# bytes are the high 8-bits of four values, and the 5th byte contains
# the packed low 2-bits of the preceding four values. In other words,
# the bits of the values A, B, C, D and arranged like so:
#
#  byte 1   byte 2   byte 3   byte 4   byte 5
# AAAAAAAA BBBBBBBB CCCCCCCC DDDDDDDD AABBCCDD
#
# Here, we convert our data into a 16-bit array, shift all values left
# by 2-bits and unpack the low-order bits from every 5th byte in each
# row, then remove the columns containing the packed bits

data = data.astype(np.uint16) << 2
for byte in range(4):
    data[:, byte::5] |= ((data[:, 4::5] >> ((4 - byte) * 2)) & 0b11)
data = np.delete(data, np.s_[4::5], 1)

# Now to split the data up into its red, green, and blue components. The
# Bayer pattern of the OV5647 sensor is BGGR. In other words the first
# row contains alternating green/blue elements, the second row contains
# alternating red/green elements, and so on as illustrated below:
#
# GBGBGBGBGBGBGB
# RGRGRGRGRGRGRG
# GBGBGBGBGBGBGB
# RGRGRGRGRGRGRG
#
# Please note that if you use vflip or hflip to change the orientation
# of the capture, you must flip the Bayer pattern accordingly

rgb = np.zeros(data.shape + (3,), dtype=data.dtype)
rgb[1::2, 0::2, 0] = data[1::2, 0::2] # Red
rgb[0::2, 0::2, 1] = data[0::2, 0::2] # Green
rgb[1::2, 1::2, 1] = data[1::2, 1::2] # Green
rgb[0::2, 1::2, 2] = data[0::2, 1::2] # Blue

# At this point we now have the raw Bayer data with the correct values
# and colors but the data still requires de-mosaicing and
# post-processing. If you wish to do this yourself, end the script here!
#
# Below we present a fairly naive de-mosaic method that simply
# calculates the weighted average of a pixel based on the pixels
# surrounding it. The weighting is provided by a byte representation of
# the Bayer filter which we construct first:

bayer = np.zeros(rgb.shape, dtype=np.uint8)
bayer[1::2, 0::2, 0] = 1 # Red
bayer[0::2, 0::2, 1] = 1 # Green
bayer[1::2, 1::2, 1] = 1 # Green
bayer[0::2, 1::2, 2] = 1 # Blue

# Allocate an array to hold our output with the same shape as the input
# data. After this we define the size of window that will be used to
# calculate each weighted average (3x3). Then we pad out the rgb and
# bayer arrays, adding blank pixels at their edges to compensate for the
# size of the window when calculating averages for edge pixels.

output = np.empty(rgb.shape, dtype=rgb.dtype)
window = (3, 3)
borders = (window[0] - 1, window[1] - 1)
border = (borders[0] // 2, borders[1] // 2)

rgb_pad = np.zeros((
    rgb.shape[0] + borders[0],
    rgb.shape[1] + borders[1],
    rgb.shape[2]), dtype=rgb.dtype)
rgb_pad[
    border[0]:rgb_pad.shape[0] - border[0],
    border[1]:rgb_pad.shape[1] - border[1],
    :] = rgb
rgb = rgb_pad

bayer_pad = np.zeros((
    bayer.shape[0] + borders[0],
    bayer.shape[1] + borders[1],
    bayer.shape[2]), dtype=bayer.dtype)
bayer_pad[
    border[0]:bayer_pad.shape[0] - border[0],
    border[1]:bayer_pad.shape[1] - border[1],
    :] = bayer
bayer = bayer_pad

# In numpy >=1.7.0 just use np.pad (version in Raspbian is 1.6.2 at the
# time of writing...)
#
#rgb = np.pad(rgb, [
#    (border[0], border[0]),
#    (border[1], border[1]),
#    (0, 0),
#    ], 'constant')
#bayer = np.pad(bayer, [
#    (border[0], border[0]),
#    (border[1], border[1]),
#    (0, 0),
#    ], 'constant')

# For each plane in the RGB data, we use a nifty numpy trick
# (as_strided) to construct a view over the plane of 3x3 matrices. We do
# the same for the bayer array, then use Einstein summation on each
# (np.sum is simpler, but copies the data so it's slower), and divide
# the results to get our weighted average:

for plane in range(3):
    p = rgb[..., plane]
    b = bayer[..., plane]
    pview = as_strided(p, shape=(
        p.shape[0] - borders[0],
        p.shape[1] - borders[1]) + window, strides=p.strides * 2)
    bview = as_strided(b, shape=(
        b.shape[0] - borders[0],
        b.shape[1] - borders[1]) + window, strides=b.strides * 2)
    psum = np.einsum('ijkl->ij', pview)
    bsum = np.einsum('ijkl->ij', bview)
    output[..., plane] = psum // bsum

# At this point output should contain a reasonably "normal" looking
# image, although it still won't look as good as the camera's normal
# output (as it lacks vignette compensation, AWB, etc).
#
# If you want to view this in most packages (like GIMP) you'll need to
# convert it to 8-bit RGB data. The simplest way to do this is by
# right-shifting everything by 2-bits (yes, this makes all that
# unpacking work at the start rather redundant...)

output = (output >> 2).astype(np.uint8)
with open('image.data', 'wb') as f:
    output.tofile(f)

This recipe is also encapsulated in the PiBayerArray class in the picamera.array module, which means the same can be achieved as follows:

import time
import picamera
import picamera.array
import numpy as np

with picamera.PiCamera() as camera:
    with picamera.array.PiBayerArray(camera) as stream:
        camera.capture(stream, 'jpeg', bayer=True)
        # Demosaic data and write to output (just use stream.array if you
        # want to skip the demosaic step)
        output = (stream.demosaic() >> 2).astype(np.uint8)
        with open('image.data', 'wb') as f:
            output.tofile(f)

New in version 1.3.

Changed in version 1.5: Added note about new picamera.array module.