Advancing Human Tracking in Multi Object Tracking Using Video Features and Depth Cues

Abstract

Abstract The identification of each objects uniquely in each video frame is known as the Multi Object Tracking(MOT). MOT is well- known for its potential uses in autonomous driving, human-robot interaction, and surveillance. The goal of MOT is to correctly identify each item and to maintain that identification over time in the face of occlusions, changes in appearance, and other obstacles. This range of uses demonstrates the effectiveness of MOT and encourages scientists to address knowledge gaps in MOT fields. This has greatly influenced the development of all these large-scale MOT works. The MOT has certain limitations. We try ton resolve this limitation using novel video features and depth cues integration. We explored freely available cross domain models such as human action recognition model and zero-shot depth extracting model. With that we explored use of depth cues to give additional information to improve tracking. We analyzed the performance of our approaches by using benchmark datasets and standard metrics. These outcome of our research is , we can use the existing cross domain models to improve MOT in the aspect of detection without further fine-tuning or complex tricks. In the other hand, usage of depth cues improves MOT performance in short videos by fusion with the existing appearance features. The major contribution of our research is, creating a unexplored path in solving MOT problem without complex algorithms and resource intensive training by utilizing existing cross domain knowledge.