Speed Up Optical Flow algorithm (If applicable) Python OpenCV

问题

I came across this interesting situation (Speeding up optical flow (createOptFlow_DualTVL1)) but it doesn't apply to my needs. My general problem is I want to speed up as much as possible the following code if it is applicable. Keep in mind, I want the frames to be grayscale and resize them to height = 300 while keeping aspect ratio locked. Also, I want to sample 2 frames per second from that video so I assume every video to be around 30fps. Finally, I want to use the TV-L1 optical flow algorithm. Is there a way to boost this algorithm because for a 1-minute video it takes around 3 minutes to estimate the optical flow which is too time-consuming for my needs.

Thanks in advance, Evan

import math, imutils, cv2
print ("Entering Optical Flow Module...")
        cap = cv2.VideoCapture(video_path)
        current_framerate = cap.get(5)
        ret, frame1 = cap.read()
        prvs = cv2.cvtColor(frame1,cv2.COLOR_BGR2GRAY)
        prvs = imutils.resize(prvs, height = 300)


        all_frames_flow=list()
        while(cap.isOpened()):
            frameId = cap.get(1)
            ret, frame2 = cap.read()
            if ret == True:
                if (frameId % (math.floor(current_framerate)/2)==0): # assume videos are 30 fps and we want only 2 frames per second.
                    next = cv2.cvtColor(frame2,cv2.COLOR_BGR2GRAY)
                    next = imutils.resize(next, height = 300)
                    optical_flow = cv2.DualTVL1OpticalFlow_create()
                    flow = optical_flow.calc(prvs, next, None)
                    all_frames_flow.append(flow)
                    prvs = next
                else:
                    continue
            else:
                break
        cap.release()

回答1:

For cv2 version "'4.1.0'":

Code below is faster but is less accurate as per the explanation of hyperparameters below. Tune these parameters to solve speed vs accuracy trade-off as per requirement.

optical_flow= cv2.optflow.DualTVL1OpticalFlow_create(nscales=1,epsilon=0.05,warps=1)
flow = optical_flow.calc(new_prvs, new_nxt, None)

• int "nscales" : Number of scales used to create the pyramid of images.

• int "warps" : Number of warpings per scale. Represents the number of times that I1(x+u0) and grad( I1(x+u0) ) are computed per scale. This is a parameter that assures the stability of the method. It also affects the running time, so it is a compromise between speed and accuracy.

• double epsilon : Stopping criterion threshold used in the numerical scheme, which is a trade-off between precision and running time. A small value will yield more accurate solutions at the expense of slower convergence.

other parameters to tune can be found here

来源：https://stackoverflow.com/questions/53209087/speed-up-optical-flow-algorithm-if-applicable-python-opencv