Abstract: Monocular pedestrian detection usually suffers from occlusion problems in complex and crowded scenes, which can lead to serious false positives. Multi-view pedestrian detection can effectively solve the occlusion problem by combining data from multiple views. In the previous multi-view detection algorithms, only single-level feature maps are used, which cannot detect multi-scale targets well. In this paper we propose a new multi-view detection algorithm which uses a newly introduced Dilated Encoder method to aggregate the information of multiple views. Dilated Encoder is a method that uses different dilated convolutions of the expansion rate so that a single layer of features gets different scale perceptual fields, covering all scale ranges of the target and improving the capability of multi-scale targets. Our proposed method achieves 90.7% MODA on the Wildtrack dataset, which is a very strong competitive result compared to the current state-of-the-art algorithms.

Key words: multi-view data, feature fusion, dilated convolution, crowded scene, pedestrian detection, multi-scale detection