Rapid discriminative detection for smart camera applications

Abstract

Tracking-by-detection is an attractive paradigm for intelligent visual surveillance applications where clutter, lighting variations, target overlap and occlusions hamper conventional background modeling. However, state-of-the-art vehicle and pedestrian detectors based on discriminative classification are too computationally expensive for real-time implementation on embedded smart cameras. This paper presents the Generative Focus of Attention-Discriminative Validation (GFA-DV) detector which uses generative target detection to greatly improve the efficiency of discriminative classification. The proposed method gains further efficiency by using a hierarchical visual codebook to enable each stage of the detector to efficiently utilize the same features within a different quantization of the feature space. This approach reduces the expense of feature matching compared to multiple flat codebooks. The proposed GFA-DV detector is experimentally compared to several state-of-the-art methods, and shown to perform better than other efficient detectors while achieving a 100 times speedup over more accurate detectors.