Real-time moving pedestrian detection using contour features

Pedestrian detection is one of the most fundamental research in computer vision. However, many high performance detectors run slowly. In this paper, we propose a real-time moving pedestrian detector by using efficient contour features. Firstly, the moving targets are detected by background subtraction. By combining the elliptic Fourier descriptors and the normalized central moments, we propose the Elliptic Fourier and Moments Descriptors (EFMD) to describe the moving target contours. Secondly, the moving targets are classified by the trained Support Vector Machine (SVM). In addition, we introduce a novel overlap handling algorithm based on linear fitting and normalized central moments, which improves the detection performance by reducing both false positives and miss rate. The experimental results on PETS 2009 and CAVIAR datasets show that our approach achieves a miss rate of 14% (PETS 2009) and 13% (CAVIAR) at 10−1 False Positives Per Image (FPPI) and an average runtime per frame of 30 ms (PETS 2009) and 25 ms (CAVIAR), which significantly outperforms several state-of-the-art detectors in both detection performance and runtime.