Abstract

This paper presents a new approach to the tracking of multiple objects in CCTV surveillance using a combination of simple neural cost functions based on Self-Organizing Maps, and a greedy assignment algorithm. Using a reference standard data set and an exhaustive search algorithm for benchmarking, we show that the cost function plays the most significant role in realizing high levels of performance. The neural cost function’s context-sensitive treatment of appearance, change of appearance and trajectory yield better tracking than a simple, explicitly designed cost function. The algorithm matches 98.8% of objects to within 15 pixels.