Neighboring and Non-Neighboring Features for Pedestrian Detection

摘要

Pedestrian detection is used in many real-time applications. However, most existing pedestrian detectors cannot achieve a tradeoff between speed and accuracy. Based on the hand-crafted feature method, this paper proposes novel non-neighboring features. The features can be used to identify pedestrians with different postures through the comparison of pedestrians' both sides. In addition, we propose corresponding neighboring features and an novel algorithm that can perform occlusion analysis. Using HOG+LUV as low-level features input and sliding window approach, this paper uses neighboring and non-neighboring features to achieve best performance on the Caltech pedestrian dataset after adding the occlusion analysis algorithm. Experimental results show that proposed model outperforms the state-of-art where it reduces the miss rate by 0.92%. In addition, the processing speed is 18 FPS in CPU environment.