Contextual person detection in multi-modal outdoor surveillance

In this paper we present a new approach to person detection in outdoor surveillance tasks. A multi-modal segmentation (RGB, Polarimetric, thermal sensors) of the world into regions sky, road, bush, trees, grass etc. is used to learn the normal spatial context of people appearing in normal training data. The context feature is a novel application of the work of Wolf et al. [1] which enables the probability of a person appearing in a certain location to be computed. By using motion as a precursor to the deployment of a HOG person detector in conjunction with the spatial context likelihood we obtain significant improvement in person detection for challenging scenes. Comprehensive ROC analysis on 4 outdoor scenes is reported for normal activity detection. Anomaly detection is then achieved using learned context and we show that 72% of true positive anomalies are found for a false positive rate of 0.19% over all data in thermal and visual band data.