Multi-Camera Tracking of Intelligent Targets with Hidden Reciprocal Chains

Real world targets are intelligent and almost always move with a destination in mind. This paper introduces a new target tracking algorithm for multi-camera networks based on a hidden reciprocal chain (HRC), which is able to capture the local dynamics and intention of a real world target in a statistical way. The model is non-causal and therefore fundamentally different to standard Markovian motion models which underpin most trackers, such as the Kalman filter. However it is less computationally expensive than more sophisticated models like Markov decision processes, which can capture complex behaviours but require approximate algorithms for inference. We argue that HRCs are a natural extension to existing Markovian models by presenting exact online inference and detection algorithms which scale well with the number of cameras and targets. Finally we demonstrate the potential benefits by presenting results on synthetic data for the problem of multi-target tracking across multiple cameras.