Motion Distillation for Pedestrian Surveillance

Motion detection in video is a fundamental step in all object tracking tasks; however, it is often incorrectly treated as a solved problem. The most common approaches are statistical background modelling and condensation. However, these methods have certain inefficiencies in their use of motion information. The output of background modelling is change detection rather than true motion detection, and is highly susceptible to non-motion pixel change due to noise and lighting changes. Condensation techniques use only spatial or ‘foreground’ information and must be initialised with target appearance information. Here we present a new wavelet-based method for true motion detection which incorporates spatial and temporal information on an equal footing. The algorithm is tested using noisy outdoor tracking scenarios, can self-initialise and track arbitrary objects, and is shown to be inherently both fast and robust.