GenTrack3: Hybrid Stochastic–Deterministic Online Multi-Object Tracking with Cluster-Aware Association

Research Approach

GenTrack3 addresses the challenge of maintaining consistent object identities in complex, real-world environments characterized by non-linear motion, occlusions, and varying target counts. The authors propose a hybrid framework that integrates stochastic particle filtering—to handle uncertainty and non-linear dynamics—with deterministic data association to ensure reproducible and efficient tracking. Unlike traditional particle filters that struggle with high-dimensional state spaces, GenTrack3 represents each target with a dedicated particle set, optimized via Markov Chain Monte Carlo (MCMC) and Particle Swarm Optimization (PSO).

Scalability and Association

A primary contribution of this work is the novel track-to-detection matching pipeline. To overcome the computational bottleneck of global cost matrices, the authors introduce a cluster-aware association strategy. By mapping tracks and detections onto a 2D occupancy grid and using connected-component labeling, the global matching problem is decomposed into smaller, independent sub-problems. This approach allows the system to handle a larger number of targets efficiently while maintaining the ability to manage group tracking and resolve ambiguities in crowded scenes.

Tracking Pipeline

The tracking process follows a systematic pipeline:

1. Particle Optimization: Particles are sampled and refined using PSO, incorporating motion consistency, appearance similarity, and social fitness (to avoid identity switches near other objects).
2. Clustering: Tracks and detections are grouped based on spatial proximity in an occupancy grid.
3. Association: A localized cost matrix is constructed for each cluster, and the Hungarian algorithm is applied to perform data association.
4. Update: Tracks are updated based on their status (strong or weak), with weak tracks receiving special handling to maintain continuity during temporary occlusions or noise.