Data Association Methods for Tracking Multiple Objects in Videos

Abstract

Tracking multiple objects in videos which are recorded by multiple non-overlapping cameras can be accomplished by resolving two sub-problems. Specifically, one is tracking objects within a camera (multi-object tracking) and the other is tracking objects across cameras (multi-camera tracking). In this dissertation, two novel data association algorithms are proposed, each of which is designed to perform a specific tracking task. For the multi-object tracking, a deep neural network architecture is proposed to solve a problem of data association between existing tracks and newly received detections at each frame. The proposed network learns the data association task from data and then multi-object tracking is accomplished as the network classifies new detections into existing objects. Second, a multiple hypothesis tracking algorithm is proposed to track objects across multiple non-overlapping cameras (multi-camera tracking). The algorithm forms track-hypothesis trees by appending a child node which indicates newly observed track. Then, a set of multi-camera trajectories is computed among all branches of track-hypothesis trees. The proposed methods are tested on challenging datasets, such as Stanford Drone Dataset and MOTchallenge 2015 2D dataset (for multi-object tracking), NLPR_MCT and DukeMTMC dataset (for multi-camera tracking). In experiments, the results demonstrate the proposed methods' strong points in comparison with other state-of-the-art methods.