Multi-Object Tracking as Continuous Dynamic Environment and its Exploration by Particle Swarm Optimization

Swarm Intelligence (SI) algorithms for Dynamic Optimization Problems (DOPs) have acquired the ability to search in environments with various dynamic changes through diverse studies. However, continuous change, one characteristic of DOPs, still needs to be thoroughly discussed. Some environments in real-world applications can be assumed to change continuously over time or very frequently. It is essential to analyze issues by actually applying the algorithm to real problems with continuous change. Multiple Object Tracking (MOT), particularly MOT via Simple Online and Real-time Tracking (SORT) can be regarded as DOP with continuous change. Estimating the next bounding box involved in SORT needs to track object trajectories as continuous change. Applying SI algorithms can expect to track nonlinear trajectories that the original SORT can not in this estimation. This paper experiments to verify the tracking performance improvement by Particle Swarm Optimization (PSO), one of the SI algorithms. The experiment employs PSO-JOFB, a variant of PSO designed for tracking continuous change. Applying PSO-JOFB can expect to confirm that exploring this problem as DOPs with continuous change can lead to more effective tracking. The dataset utilized for the validation is the MOT Challenge 2015 training data in this experiment. The experimental results reveal the following implications: (1) The number of evaluations can be significantly reduced by exploring the next bounding box as DOPs with continuous change rather than as a static environment; (2) Tracking by PSO and PSO-JOFB enables the tracking trajectories that the original SORT fails to track.