Multi-Object Tracking (MOT) is a fundamental task in computer vision, vital for applications in autonomous driving, intelligent surveillance, and sports data analysis. However, tracking performance significantly degrades under conditions such as occlusion, small object instances, and fast motion. This work proposes TraceNet, a modular multi-object tracking framework designed to address these challenges by incorporating sophisticated detection, association, and recovery components. TraceNet builds on a fine-tuned YOLOv11 detector and incorporates a Confidence Optimization Network (CON) to improve detection reliability in low-visibility environments. It further includes a Deep Similarity Integration (DSI) module improved by Dynamic IoU Adjustment (DIA), which combines motion prediction and appearance cues to achieve reliable identification associations. The framework uses a Graph-Based Track Recovery (GBTR) network and a Neural Trajectory Smoother (NTS) to recover interrupted trajectories and ensure temporal consistency. The temporal association is further enhanced by a Transformer-Based Association (TBA) module. TraceNet achieves exceptional performance on four challenging benchmarks, achieving HOTA scores of 66.9 for MOT17 and 66.7 for MOT20, with IDF1 scores of 83.2 and 83.5, respectively. These results highlight TraceNet’s robustness in dense and occluded scenes, and demonstrate that it is a high-performing and scalable solution for real-time multi-object tracking. • Detection Enhancement: Uses a fine-tuned YOLOv11 together with the Confidence Optimization Network (CON) to improve accuracy for small, occluded, and fast-moving objects. • Deep Similarity Integration (DSI) with Dynamic Adjustment: proposes a Siamese neural network-based DSI module, consisting of an Occlusion-Handling Network (OHN), a Motion Consistency Module (MCM), and a Shape Matching Network (SMN), integrated with Dynamic IoU Adjustment (DIA) to compute a robust similarity metric that reduces identity switches in occluded and crowded conditions. • Graph-Based Track Recovery (GBTR): utilizes a Graph Neural Network (GNN) within the GBTR module to model inter-object relationships, enabling effective recovery of lost tracks in complex situations. • Modular Real-Time Optimization : Integrates a Transformer-Based Association (TBA) and Neural Trajectory Smoother (NTS) into a modular architecture, optimizing detection, association, and trajectory smoothing for real-time performance.
Khan et al. (Thu,) studied this question.