SpaFIND: an Effective and Low-cost Feature Descriptor for Pedestrian Protection Systems in Economy Cars

Abstract

Pedestrian protection systems (PPSs) are crucial for reducing road traffic fatalities. However, the relatively high cost of today's PPSs prevents the majority of economy cars from receiving their benefits. In this paper, we propose an effective and low-cost sparse feature interaction descriptor (SpaFIND) that is designed for real-time pedestrian detection with limited computational power in economy cars. SpaFIND extends the histogram of oriented gradients (HOG) feature and selectively computes the pairwise relationships between neighboring components of the HOG. Therefore, SpaFIND can capture the second-order properties of object appearance while maintaining a low computational load, which is the main contribution of this paper. During the experiments, we performed comparative evaluations of the proposed system against several baseline methods on the Caltech Pedestrian Detection Benchmark in terms of detection accuracy and computational load. The results demonstrated the effectiveness of the proposed low-cost SpaFIND feature, which can meet the requirements for implementing a PPS with limited computational power, thereby contributing to the massive deployment of PPSs in economy cars in the immediate future.