view:42069 Last Update: 2023-11-16
Mahnaz Moghaddam, Mostafa Charmi, Hossein Hassanpoor
A robust attribute-aware & real-time multi-target multi-camera tracking system using multi-scale enriched features and hierarchical clustering
Multi-Camera Multi-Target Tracking (MTMCT) has challenges such as viewpoint and pose variations, scale and illumination changes, and occlusion. Available MTMCT approaches have high computational complexity and are not sufficiently robust in the mentioned challenges. In this work, an Attribute Recognition-based MTMCT(AR-MTMCT) framework is presented for real-time application. This framework performs object detection, re-Identification (Re-Id) feature extraction, and attribute recognition in an end-to-end manner. Applying attributes highly improves MTMC online tracking performance in the mentioned challenges. The pipeline of AR-MTMCT consists of three modules. The first module is a novel one-shot Single-Camera Tracking (SCT) architecture named Attribute Recognition-Multi Object Tracking (AR-MOT) which performs object detection, Re-Id feature extraction, and attributes recognition using one backbone through multi-task learning. Hierarchical clustering is performed in the second module to deal with the detection of several instances of one identity in the overlapping areas of cameras. In the last module, a new data association algorithm is performed using spatial information to reduce matching candidates. We also have proposed an efficient strategy in the data association algorithm to remove lost tracks by making a trade-off between the number of lost tracks and the maximum lost time. Evaluation and training of AR-MTMCT have been done on the large-scale MTA dataset. The proposed system has been improved by 20% and 11%, respectively, compared to the WDA method in IDF1 and IDs metrics. Also, the AR-MTMCT outperforms the state-of-the-art methods by a large margin on decreasing computational complexities.