Pursuing Detector Efficiency for Simple Scene Pedestrian Detection

Detector accuracy is by any means the key focus in most existing pedestrian detection algorithms especially for clutter scenes. However, it is not always necessary, while sometimes over-fitted, to directly leverage such detectors in scenarios with simple scene compositions. To this end, limited work has done on a systematic detector simplification towards balancing its speed and accuracy. In this paper, we study this problem by investigating two mutually correlated issues, i.e. fast edge-based feature extraction and detector score computation. For handling the first issue, a simple Structured Local Edge Pattern (SLEP) is proposed to extract and encode local edge cues, extremely effectively, into a histogram. For the second, an integral image based acceleration is proposed toward fast classifier score computation by transforming the classifier score into a linear sum of weights. Experimental results on CASIA gait recognition dataset show that our proposed method is highly efficient than most existing detectors, which even faster than the practical OpenCV pedestrian detector.