Human Tracking for Facility Surveillance

Abstract

This research provides two main changes based on Detect-And-Track. To improve the Multi-Object Tracking Accuracy (MOTA) while keeping the lightweight of the original approach, this paper proposes a gradient approach to obtain higher MOTA. We use the location of two previous frames of the same identified person to calculate the gradient for the location prediction of the current frame. Then, the predicted and the detected locations are compared. We also compare the current and previous detections. With a weighted combination for matching, we increase the MOTA score and improve the results of Detect-And-Track. Moreover, this research replaces cosine distance, the original feature extractor, with Euclidean distance. By doing so, feature extraction can match Intersection over Union (IoU) better. The weighted combination, which consists of IoU and Euclidean distance, provides a better MOTA than Detect-And-Track. In addition, a greedy approach facilitates a higher MOTA when implement with IoU and Euclidean distance. This weighted combination utility is superior than the combination of IoU and cosine distance, achieving 56.1% MOTA in total on the validation data of PoseTrack ICCV’17 dataset.