A KNN Match Based Tracking-Learning-Detection Method with Adjustment of Surveyed Areas

Abstract

A Tracking-Learning-Detection (TLD) is a tracking framework that decomposes the long-term tracking task into tracking, learning and detection. To improve the tracking accuracy and precision with TLD, the constraint that a target cannot move a significant distance within a short time period with limited velocity is added, and the target displacement in the previous frame analyzed is used to reduce the detection range of the TLD detector. Instead of using the Median Flow Tracker, a K-Nearest Neighbor (Knn) matching with the Fused Lukas-Kanade method with ORB feature points is used to track targets. Experiments demonstrate that the improved TLD algorithm has better tracking accuracy.