Dempster-Shafer Multifeature Fusion for Pedestrian Detection

Abstract

Pedestrian detection is of great importance for ensuring traffic safety. In recent years, many works employing image-based shape features to recognize pedestrians have been reported. However, previous pedestrian detectors were in many cases not sufficient to achieve satisfactory results under complex weather conditions and complex scenarios. As a solution this paper exploits two video-based motion feature descriptors and applies such motion features to the detection task in addition to four classical shape features with the aim of significantly improving the detection performance. Our motion features are defined as the trajectory smoothness degree and motion vector field, which are derived from our proposed point tracking strategy beyond tough target segmentation. And then the appealing Dempster-Shafer theory of evidence (D-S theory) is applied to fuse these features, due to the fact that D-S theory is better than the classical Bayesian approach in handling the information with lack of prior probabilities. The proposed automatic pedestrian detection algorithm is evaluated on real data and in real traffic scenes under various weather conditions. Theoretical analysis and experiment results consistently show that the proposed method outperforms SVM-based multifeature fusion approach for pedestrian detection in terms of recognition ability and robustness in various real traffic scenes.