Gaining high frame rate from ELP camera (python)
I'm using an ELP USB camera that is supposedly rated at 100fps at 640x480 quality but I don't seem to be getting anywhere near that frame rate. I was wondering if anyone has used an ELP camera with openCV and managed to achieve a frame rate close to 100 fps? Or if anyone has any advice to help increase the frame rate? I am using Ubuntu 14.04 with openCV version 2.4.8.
I have already tested that the output is MJPEG at 640x420 and I am running cv2.VideoCapture in one thread and placing the frame on a queue. From the main thread I repeatedly ask for frames but only return when the results is not None. This gives a frame rate hovering around 30fps. One I have the image I am performing contour detection but essentially I would like to get the frame rate up as high as possible. Below is the code I'm using just to test the frame rate.
import cv2
import time
import numpy as np
from datetime import datetime
from threading import Thread, Lock, Condition
import time
from Queue import Queue
class WebcamVideoStream:
def __init__(self, src=0):
# initialize the video camera stream
self.stream = cv2.VideoCapture(src)
# initialize the variable used to indicate if the thread should
# be stopped
self.stopped = False
self.frame = None
def start(self):
global qt
self.stopped = False
qt = Queue(10)
# start the thread to read frames from the video stream
thread1 = Thread(target=self.update, args=())
thread1.start()
return self
def update(self):
global qt
# keep looping infinitely until the thread is stopped
while True:
if self.stopped:
return
_, self.frame = self.stream.read()
qt.put(self.frame)
def read(self):
global qt
if(not qt.empty()):
self.CurrFrame=qt.get()
if self.CurrFrame is not None:
return self.CurrFrame
if self.stopped:
return
def stop(self):
print('Stop')
# indicate that the thread should be stopped
self.stopped = True
return self
vs = WebcamVideoStream(src=-1).start()
time.sleep(1)
i = 0
t0 = time.time()
while i < 100:
frame = vs.read()
while frame is None:
frame = vs.read()
i = i + 1
rate = 100/(time.time()-t0)
print(rate)
cv2.destroyAllWindows()
vs.stop()
Thanks in advance
I'm struggling to get the embedded code to format, sorry.
os ? opencv version ?
(i.e. on win, dshow will force that into bgr, which takes time, but what would you want to do even with a mjpeg image, if not uncompress it ?)
then, this is basically an io bottleneck, you won't gain anything with multithreading
In the demo I'm doing absolutely nothing with them. I had assumed that because (most of the time) when a frame is requested None is returned that the requests are coming in far faster than the camera can output them and therefore the queue never has more than one frame on it.
apologies, i removed a comment, you the other, so the missing information:
again, i'm quite sure, that cv2.VideoCapture will have to uncompress your mjpg image in read()
(else you could not use it for contour detection)
Ah ok. So this is causing the bottleneck? Is there any method to use to improve the frame rate?
i don't think, there's much you can do from python . if it was c++, i'd say: try to use libv4l directly, without using VideoCapture (there's also a hidden fifo queue, threads & locks & whatnot in the v4l wrapper)
I decided to upgrade to OpenCV 3 and this has made all the difference. Frame rate immediately rose greatly. Thanks for all your help, Berak.