A Novel Descriptor for Pedestrian Detection Based on Multi-layer Feature Fusion

Pedestrian detection, as a research focus of computer vision, is effectively utilized in the fields of intelligent security and traffic. The puzzle of pedestrian detection is scene's complexity, pedestrian's multi-pose and pedestrian occlusion. Furthermore, other issues need to be considered in practical applications, such as environment illumination and humidity. Therefore, performance is required to be raised in aspects, such as accuracy, robustness, and velocity of detection algorithm. In this paper, Histogram of Oriented Gradients (HOG) and multi-layer (set to 3) Local Binary Patterns (LBP) features are concatenated in sequence to form a novel type of multi-layer feature. Then the fusion features are classified by SVM. Experiments and results confirm the feasibility of the proposed method.