Multiple human tracking in wearable camera videos with informationless intervals

Multiple human tracking plays a key role in video surveillance and human activity detection. Compared to fixed cameras, wearable cameras, such as GoPro and Google Glass, can follow and capture the targets (people of interest) in larger areas and from better view angles, following the motion of camera wearers. However, wearable camera videos suffer from sudden view changes, resulting in informationless (temporal) intervals of target loss, which make multiple human tracking a much more challenging problem. In particular, given large and unknown camera-pose change, it is difficult to associate the multiple targets over such an interval based on the spatial proximity or appearance matching. In this paper, we propose a new approach, where spatial pattern of the multiple targets are extracted, predicted and then leveraged to help associate the targets over an informationless interval. We also propose a classification based algorithm to identify the informationless intervals from wearable camera videos. Experiments are conducted on a new dataset containing 30 wearable-camera videos and the performance is compared to several other multi-target tracking algorithms.