OpenCV CAMshift跟踪：目标移出画面/质量过低时如何终止跟踪？

Great question! CAMshift itself doesn’t come with built-in features for stopping when a target leaves the frame or terminating based on tracking quality, but you can easily implement both checks yourself with a bit of extra code. Let’s break this down step by step:

1. Stopping Tracking When the Target Moves Off-Screen #

You can manually verify if the tracked object’s bounding area falls outside the frame boundaries. Here’s how to put this into practice:

First, grab your video frame dimensions, then extract the rotated bounding rectangle returned by CAMshift. You can either check if the rectangle’s center is outside the frame (a simpler check) or verify that all its vertices are off-screen (a stricter check)—whichever fits your use case better.

Example code snippet:

import cv2
import numpy as np

# Initialize CAMshift setup
cap = cv2.VideoCapture(0)
ret, frame = cap.read()
# Select initial ROI (you can replace this with a predefined region if needed)
x, y, w, h = cv2.selectROI(frame, False)
track_window = (x, y, w, h)

# Create target histogram
roi = frame[y:y+h, x:x+w]
hsv_roi = cv2.cvtColor(roi, cv2.COLOR_BGR2HSV)
mask = cv2.inRange(hsv_roi, np.array((0., 60., 32.)), np.array((180., 255., 255.)))
roi_hist = cv2.calcHist([hsv_roi], [0], mask, [180], [0, 180])
cv2.normalize(roi_hist, roi_hist, 0, 255, cv2.NORM_MINMAX)
term_crit = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, 10, 1)

while True:
    ret, frame = cap.read()
    if not ret:
        break
    frame_h, frame_w = frame.shape[:2]
    hsv = cv2.cvtColor(frame, cv2.COLOR_BGR2HSV)
    dst = cv2.calcBackProject([hsv], [0], roi_hist, [0, 180], 1)
    
    # Run CAMshift
    ret, track_window = cv2.CamShift(dst, track_window, term_crit)
    pts = cv2.boxPoints(ret)
    pts = np.int0(pts)
    
    # Check if target is off-screen (using center point check)
    center_x = int((pts[0][0] + pts[2][0]) / 2)
    center_y = int((pts[0][1] + pts[2][1]) / 2)
    off_screen = center_x < 0 or center_x > frame_w or center_y < 0 or center_y > frame_h
    
    if off_screen:
        print("Target has left the frame. Stopping tracking.")
        break
    
    # Draw tracking rectangle
    cv2.polylines(frame, [pts], True, (0, 255, 0), 2)
    cv2.imshow('CAMshift Tracking', frame)
    
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

cap.release()
cv2.destroyAllWindows()

2. Terminating Tracking Based on Low Quality #

CAMshift relies on back-projection of the target’s histogram, so you can use metrics tied to this back-projection to assess tracking reliability. Two common approaches are:

• Check the peak value in the back-projection window: When the target is lost, the maximum value in the tracking window’s back-projection will drop sharply.
• Compare current target histogram with the initial one: If the similarity falls below a threshold, the target is likely no longer being tracked accurately.

Example code to add quality checks to your loop:

# Inside the while loop, after getting track_window
x, y, w, h = track_window

# Option 1: Check back-projection peak value
roi_dst = dst[y:y+h, x:x+w]
max_val = np.max(roi_dst)
# Adjust threshold based on your setup (start with 50-100 and tweak)
if max_val < 60:
    print("Tracking quality too low. Stopping.")
    break

# Option 2: Compare current ROI histogram with initial
current_roi = frame[y:y+h, x:x+w]
current_hsv = cv2.cvtColor(current_roi, cv2.COLOR_BGR2HSV)
current_mask = cv2.inRange(current_hsv, np.array((0., 60., 32.)), np.array((180., 255., 255.)))
current_hist = cv2.calcHist([current_hsv], [0], current_mask, [180], [0, 180])
cv2.normalize(current_hist, current_hist, 0, 255, cv2.NORM_MINMAX)
# Use correlation (higher = better; threshold around 0.3-0.5 works for most cases)
similarity = cv2.compareHist(roi_hist, current_hist, cv2.HISTCMP_CORREL)
if similarity < 0.4:
    print("Target histogram mismatch. Stopping tracking.")
    break

You can combine both checks (off-screen detection + quality threshold) for more robust termination. Just adjust the thresholds based on your specific video and target to get the best results.

内容的提问来源于stack exchange，提问作者oka96