Low complexity multi-target tracking for embedded systems

This paper investigates the Multi-Target Tracking (MTT) for embedded systems. Many real-time tracking applications, like surveillance with Unmanned Aerial Vehicle (UAV), objects tracking with smart phones or smart cameras, require the use of embedded systems that are constrained by the size and the energy consumption of the calculator. However, the existing MTT methods are not compatible with these constrains: small amount of memory, relatively low frequency clock and lack of Floating Point Unit (FPU). To overcome the challenge of using MTT algorithms in embedded platforms, we first compared the existing MTT methods according to their tracking quality and processing complexity. We selected Gaussian Mixture Probability Hypothesis Density filter (GM-PHD) for an embedded implementation. To decrease the complexity of the algorithm and to ensure the realtime processing, we performed some approximations to simplify its operations and floating-point to fixed-point conversion to improve its performance on an embedded processor without FPU. The results obtained demonstrate the possibility for MTT to be implemented in an embedded system with almost no degradation of the tracking quality and the overall performances.