Search code examples
pythonlinuxmultithreadingraspberry-pijpeg

Raspberry Pi 3 Increasing FPS with 720p USB Camera


I write some code in python to open USB camera and grab frame from it. I use my code for http stream. For JPEG encode I use libturbojpeg library. For this I use 64 bit OS.

product: Raspberry Pi 3 Model B Rev 1.2
serial: 00000000f9307746
width: 64 bits
capabilities: smp cp15_barrier setend swp

I do some test with different resolutions.

Resolution   FPS   Time for encode
640 x 480     ~35       ~0.01
1280 x 720    ~17       ~0.028

And this is my code

import time
import os
import re
import uvc
from turbojpeg import TurboJPEG, TJPF_GRAY, TJSAMP_GRAY
jpeg = TurboJPEG("/opt/libjpeg-turbo/lib64/libturbojpeg.so")
camera = None

import numpy as np
from threading import Thread

class ProcessJPG(Thread):

    def __init__(self, data):
        self.jpeg_data = None
        self.data = data
        super(ProcessJPG, self).__init__()

    def run(self):
        self.jpeg_data = jpeg.encode((self.data))

    def get_frame(self):
        self.frame = camera.get_frame()

global camera
dev_list = uvc.device_list()
print("devices: ", dev_list)
camera = uvc.Capture(dev_list[1]['uid'])
camera.frame_size = camera.frame_sizes[2] // set 1280 x 720
camera.frame_rate = camera.frame_rates[0] // set 30 fps

class GetFrame(Thread):
    def __init__(self):
        self.frame = None
        super(GetFrame, self).__init__()
    def run(self):
        self.frame = camera.get_frame()

_fps = -1
count_to_fps = 0
_real_fps = 0
from time import time
_real_fps = ""
cfps_time = time()

while True:
    if camera:
        t = GetFrame()
        t.start()
        t.join()
        img = t.frame
        timestamp = img.timestamp
        img = img.img
        ret = 1
    t_start = time()
    t = ProcessJPG(img)
    t.start()
    t.join()
    jpg = t.jpeg_data
    t_end = time()
    print(t_end - t_start)
    count_to_fps += 1
    if count_to_fps >= _fps:
        t_to_fps = time() - cfps_time
        _real_fps = 1.0 / t_to_fps
        cfps_time = time()
        count_to_fps = 0
    print("FPS, ", _real_fps)

Encoding line is: jpeg.encode((self.data))

My question is, it is possible to increase FPS for 1280 x 720 (eg 30fps) resolution or should I use more powerful device? When I look on htop during the computation CPU is not used in 100%.

EDIT: Camera formats:

[video4linux2,v4l2 @ 0xa705c0] Raw       :     yuyv422 :           YUYV 4:2:2 : 640x480 1280x720 960x544 800x448 640x360 424x240 352x288 320x240 800x600 176x144 160x120 1280x800
[video4linux2,v4l2 @ 0xa705c0] Compressed:       mjpeg :          Motion-JPEG : 640x480 1280x720 960x544 800x448 640x360 800x600 416x240 352x288 176x144 320x240 160x120

enter image description here


Solution

  • It is possible and you don't need more powerful hardware.

    From the pyuvc README.md,

    * Capture instance will always grab mjpeg conpressed frames from cameras.

    When your code accesses the .img property, that invokes jpeg2yuv (see here and here). Then you are re-encoding with jpeg_encode(). Try using frame.jpeg_buffer after the capture and don't touch .img at all.

    I took a look at pyuvc on an RPi2 with a Logitech C310 and made a simplified example,

    import uvc
    import time
    
    def main():
        dev_list = uvc.device_list()
        cap = uvc.Capture(dev_list[0]["uid"])
        cap.frame_mode = (1280, 720, 30)
        tlast = time.time()
        for x in range(100):
            frame = cap.get_frame_robust()
            jpeg = frame.jpeg_buffer
            print("%s (%d bytes)" % (type(jpeg), len(jpeg)))
            #img = frame.img
            tnow = time.time()
            print("%.3f" % (tnow - tlast))
            tlast = tnow
        cap = None
    
    main()
    

    I get ~.033s per frame, which works out to ~30fps at ~8%CPU. If I uncomment the #img = frame.img line it goes up to ~.054s/frame or ~18fps at 99%CPU (the decode time limits the capture rate).