Bilateral association tracking with parzen window density estimation

Multi-object tracking is an important branch of computer vision, which is mostly used for behavior recognition and event analysis. At present, most of the research focuses on the accuracy of tracking. However, the real-time performance is also urgently desired but there is lack of research. As a result, Deepsort, proposed 5 years ago, is still the most widely used tracker in real applications. In this paper, a Bilateral Association Tracking (BAT) framework is proposed. It uses tracklet as the basic node instead of discrete detection for tracking. Meanwhile, a Parzen density based Hierarchical Agglomerative Clustering (P-HAC) algorithm is introduced to describe the density distribution of targets and generate tracklets with high confidence. In addition, Dual Appearance Features (DAF) is proposed which considers both spatial and temporal features of tracklets and promotes the accuracy of tracklet association. Experiments are conducted on popular benchmarks such as MOT2017, Visdrone and KITTI. BAT outperforms Deepsort on both association accuracy and trajectory integrity without obvious efficiency decline. Compared with other state-of-the-art trackers, BAT shows significant advantage on computational cost while performing competitive tracking accuracy as well. It is hoped that the research can promote the applications on real-time tracking in the near future.

Automated summary

Multi-object tracking is an important branch of computer vision used for behavior recognition and event analysis. Most research currently focuses on tracking accuracy, with a lack of research on real-time performance. The proposed Bilateral Association Tracking (BAT) framework uses tracklet as the basic node for tracking and introduces a Parzen density based Hierarchical Agglomerative Clustering (P-HAC) algorithm for generating high confidence tracklets. The Dual Appearance Features (DAF) approach considers both spatial and temporal features, improving tracklet association accuracy. BAT outperforms Deepsort in association accuracy and trajectory integrity without obvious efficiency decline, showing significant advantage on computational cost compared to other state-of-the-art trackers while maintaining competitive tracking accuracy. This research aims to promote real-time tracking applications in the future.