Local sparse appearance model with specific structural information in infrared pedestrian tracking

Abstract

Robust pedestrian tracking in infrared image sequence becomes a crucial requirement for numerous computer vision applications. Conventional sparse appearance models have been widely used for tracking in infrared image sequence because it can reduce the noise efficiently and find the inner information in the fewer model data, but most of them aim at tracking the general object or rigid object and the structural information of an object didn't be exploited, so it cannot work efficiently when tracking the non-rigid object such as pedestrian. Based on ALSA, this paper proposes further an adaptive part-based local sparse appearance model for reducing the influence by the non-rigidity shapes to the pedestrian. Then, two-step pooling algorithm is presented to extract the moving characteristics of pedestrian such as walking and swing arm. A confidence level evaluation is used to adapt the drastic appearance changes and improve the template updating strategy when the particle filter is working. Both qualitative and quantitative evaluations are carried out on four infrared image sequences collected by ourselves and OSU dataset. The results demonstrate that the proposed algorithm can outperform the other start-of-the-art algorithms when occluding, scale changing and cluster background occurs.