Visual Tracking in Background Subtracted Image Sequences via Multi-Bernoulli Filtering

This correspondence presents a novel method for simultaneous tracking of multiple non-stationary targets in video. Our method operates directly on the video data and does not require any detection. We propose a multi-target likelihood function for the background-subtracted grey-scale image data, which admits multi-target conjugate priors. This allows the multi-target posterior to be efficiently propagated forward using the multi-Bernoulli filter. Our method does not need any training pattern or target templates and makes no prior assumptions about object types or object appearance. Case studies from the CAVIAR dataset show that our method can automatically track multiple targets and quickly finds targets entering or leaving the scene.