Multi-view Target Transformation for Pedestrian Detection

Abstract

Occlusion is one of the most challenging problems in single-view pedestrian detection. To alleviate the occlusion problem, multi-view systems have been exploited to fully acquire and recognize blocked targets. Most often, methods from the literature exploit perspective transformation to aggregate different sensing view angles of the scene, but projection distortion issues cause spatial structure break and prevent these methods from fully exploring the projected features. In this paper, we propose a novel approach, Multiview Target Transformation (MVTT), to address the distortion problem inherent in multi-view aggregation by encoding the full target features and limiting the area of interest of the projected features. Experiment results show that the performance of our proposed method compares favorably against recent relevant methods on public datasets. The ablation studies also confirm the effectiveness of the proposed components.