Vehicle video detection and tracking quality analysis

Abstract

This paper considers the problem of vehicle video detection and tracking. A solution based on the partitioning a video into blocks of equal length and detecting objects in the first and last frames of the block is proposed. Matching of vehicle locations in the first and last frames helps detect pairs of locations of the same object. Reconstruction of vehicle locations in the intermediate frames allows restoring separate parts of motion tracks. Combination of consecutive segments by matching makes it possible to reconstruct a complete track. Analysis of detection quality shows a true positive rate of more than 75% including partially visible vehicles, while the average number of false positives per frame is less than 0.3. The results of tracking of separate vehicles show that objects are tracked to the final frame. For the majority of them the average overlapping percent is not less efficient than the currently used Lucas-Kanade and Tracking-Learning-Detection methods. The average tracking accuracy of all vehicles makes about 70%.