[Blog] CV2 Video, and motion detection and tracking

In the previous blog posts, we explored basic image processing techniques in OpenCV.

While OpenCV is excellent for image manipulation, it also excels at processing video frames. By using cap = cv2.VideoCapture(video_file), we can analyze video streams frame by frame.

In this post, I will demonstrate motion detection and tracking techniques using a sample video of Shibuya Crossing, extracted from a public live stream. If you'd like to follow along with the same video, you can download it here:

To apply this method to the entire video:

prev_frame = cap.read()[1]

while cap.isOpened():

    ret, frame = cap.read()

    if ret:

        delta_frame = cv2.absdiff(prev_frame, frame)

        cv2.imshow('Delta Frame', delta_frame)

        prev_frame = frame

        if cv2.waitKey(1) & 0xFF == ord('q'):

            break

    else:

        break



cap.release()

cv2.destroyAllWindows()

Saving the video

This above code displayed the video but it closed after the code was finished. To save this video we can use cv2.VideoWriter(). This is quite simple all we need to do is set the recorder variable then write each frame to it in the loop, then close it when done.

rec = cv2.VideoWriter('./output_delta.mp4',cv2.VideoWriter_fourcc(*'mp4v'),fps,(frame_width,frame_height),True)





#Video test 1

prev_frame=cap.read()[1]

while(cap.isOpened()):

    ret, frame = cap.read()

    if ret == True:

        delta_frame = cv2.absdiff(prev_frame,frame)

        cv2.imshow('Delta Frame', delta_frame)

        prev_frame=frame

        rec.write(delta_frame)

        # Press "q" to quit option

        if cv2.waitKey(1) & 0xFF == ord('q'):

            break

    else:

        break





cap.release()

cv2.destroyAllWindows()

rec.release()

Note is good process to just use the fps,frame_width,frame_height, from the cap values to not lead to any issues with the size of the final video.

Another option is change the output format. Its currently set to mp4 by the cv2.VideoWriter_fourcc(*'mp4v') but over formats like mov and avi can be used too.

Method 2: Background Subtraction

Using cv2.createBackgroundSubtractorMOG2, we can separate moving objects from the static background.

fgbg = cv2.createBackgroundSubtractorMOG2(history=500, varThreshold=16, detectShadows=True)



while cap.isOpened():

    ret, frame = cap.read()

    if ret:

        fgmask = fgbg.apply(frame)

        cv2.imshow('Background Subtraction', fgmask)

        if cv2.waitKey(1) & 0xFF == ord('q'):

            break

    else:

        break



cap.release()

cv2.destroyAllWindows()

cv2.createBackgroundSubtractorMOG2(history = 500,varThreshold = 16, detectShadows = true)

• history: The number of frames used to build the background model. The default value is 500.


• ・varThreshold: The threshold on the squared Mahalanobis distance between the pixel and the model to decide whether a pixel is well described by the background model. The default value is 16.


• ・detectShadows: A Boolean indicating whether to detect shadows. If set to True, the algorithm detects and marks shadows, which are typically shown in gray in the output mask. The default value is True.

Now using this method, the motion is much brighter with enter bodies being highlighted, we can even see some differences in people who are walking across the crossing. Like someone holding an umbrella. But there is a lot more noize in this version with some background artifacts from frame to frame.

As Shadows are detected in fgbg object, a separate value is set to them than the object which is solid white. we can do this by a simple array boolean change to set all non 255 values to zero

fgmask[fgmask != 255] = 0

There is also still a lot of noize in the version . One way to reduce this is use the value channel from a hsv image then apply a slight blur to cut out those artifacts

def process_frame(frame):

    frame= cv2.cvtColor(frame, cv2.COLOR_BGR2HSV)[:,:,2]





    frame = cv2.GaussianBlur(frame, (5, 5), 1)

    return frame

This reduced the pixel artifact but does not solve them entirely but its good enough for now.

Method 3 cv2.calcOpticalFlowPyrLK

Enhancing Motion Tracking with cv2.calcOpticalFlowPyrLK

In previous sections, we explored motion detection based on changes in pixel values between frames. However, detecting motion alone doesn't provide insights into how individual objects move. OpenCV's cv2.calcOpticalFlowPyrLK is a powerful tool for tracking specific points across video frames. This function uses the Lucas-Kanade method for optical flow estimation, making it highly effective for tracking dynamic objects in scenes such as the bustling traffic at Shibuya crossing.

import cv2

import numpy as np



# Parameters for Shi-Tomasi corner detection

feature_params = dict(maxCorners=100, qualityLevel=0.3, minDistance=7, blockSize=7)



# Parameters for Lucas-Kanade optical flow

lk_params = dict(winSize=(15, 15), maxLevel=2, criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 0.03))



# Video capture and output setup

cap = cv2.VideoCapture("shibuya_traffic.mp4")

fourcc = cv2.VideoWriter_fourcc(*'mp4v')

rec = cv2.VideoWriter("output_tracking.mp4", fourcc, 30.0, (int(cap.get(3)), int(cap.get(4))))



# Read the first frame and initialize tracking points

ret, old_frame = cap.read()

old_gray = cv2.cvtColor(old_frame, cv2.COLOR_BGR2GRAY)

p0 = cv2.goodFeaturesToTrack(old_gray, mask=None, **feature_params)



while True:

    ret, frame = cap.read()

    if not ret:

        break



    # Convert the current frame to grayscale

    frame_gray = cv2.cvtColor(frame.copy(), cv2.COLOR_BGR2GRAY)



    # Calculate optical flow

    p1, st, err = cv2.calcOpticalFlowPyrLK(old_gray, frame_gray, p0, None, **lk_params)



    # Select points successfully tracked

    good_new = p1[st == 1]

    good_old = p0[st == 1]



    # Draw the tracked points

    for i, (new, old) in enumerate(zip(good_new, good_old)):

        a, b = map(int, new.ravel())

        c, d = map(int, old.ravel())

        cv2.circle(frame, (a, b), 5, (0, 0, 255), -1)

        cv2.line(frame, (a, b), (c, d), (0, 255, 0), 2)



    # Update the previous frame and points

    old_gray = frame_gray.copy()

    p0 = good_new.reshape(-1, 1, 2)



    # Write and display the frame

    rec.write(frame)

    cv2.imshow("Tracked Motion", frame)



    if cv2.waitKey(1) & 0xFF == ord('q'):

        break



cap.release()

rec.release()

cv2.destroyAllWindows()

Setting Key Points for Tracking

Before tracking can begin, initial key points must be identified. This is done using the cv2.goodFeaturesToTrack function:

p0 = cv2.goodFeaturesToTrack(old_gray, mask=None, **feature_params)

Parameters for feature_params:

• maxCorners: The maximum number of corners (keypoints) to detect. Increasing this value allows more points to be tracked but increases computational cost.


• qualityLevel: The minimum accepted quality of detected corners (0 to 1). Lower values detect weaker corners but may reduce tracking reliability.


• minDistance: Minimum Euclidean distance between detected points. Smaller values allow detection of closer points, increasing density.


• blockSize: Size of the neighborhood considered for corner detection. Smaller values detect finer details but are more sensitive to noise.

Lucas-Kanade Optical Flow Parameters

lk_params = dict(winSize=(15, 15), maxLevel=2, criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 0.03))

• 
  winSize: Defines the size of the search window for each pyramid level. Smaller windows increase precision but may miss large-scale motion.


• 
  maxLevel: Sets the number of pyramid levels used for optical flow estimation. Higher levels improve tracking over larger displacements but require more computation.


• 
  criteria: Determines the stopping condition for the iterative search algorithm:
  
  
  
  
  • EPS: Stops when the search achieves a small enough change.
  
  
  • COUNT: Stops after a fixed number of iterations.
  
  
  
  
  
  

In a complex scene like Shibuya crossing, adjusting parameters can enhance tracking results. Increasing the number of corners and reducing the search window size can improve motion precision:

# Optimized parameters

feature_params = dict(maxCorners=500, qualityLevel=0.3, minDistance=7, blockSize=7)

lk_params = dict(winSize=(7, 7), maxLevel=5, criteria=(cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 0.03))