Adaptive Structural Model for Video Based Pedestrian Detection

The performance of generic pedestrian detector usually declines seriously for videos in novel scenes, which is one of the major bottlenecks for current pedestrian detection techniques. The conventional works improve pedestrian detection in video by mining new instances from detections and adapting the detector according to the collected instances. However, when treating the two tasks separately, the detector adaptation suffers from the defective output of instance mining. In this paper, we propose to jointly handle the instance mining and detector adaption using an adaptive structural model. The regularization function of the model is applied on detector to prevent overfitting in adaption, and the loss function is designed to evaluate the combination of mined instances set and detector. Particularly, we extend the Deformable Part Model (DPM) to adaptive DPM, where an adaptive feature transformation defined on low-level HOG cell is learned to reduce the domain shift, and the regularization function for the detector is conducted on the transformation. The loss of the instance set and detector is measured by a cost-flow network structure which incorporates both the appearance of frame-wise detections and their spatio-temporal continuity. We demonstrate an alternating minimization procedure to optimize the model. The proposed method is evaluated on ETHZ, PETS2009 and Caltech datasets, and outperforms baseline DPM by 7% in terms of mean miss rate.