Pedestrian detection under occlusion scenes remains a formidable challenge in computer vision. Recently, anchor-free approach has been raised on the object detection and pedestrian detection field, anchor-free detector Center and Scale Prediction (CSP) has been proposed in pedestrian detection without special measures for occlusion. In this paper, we propose an anchor-free detector named OCSP with power ful occlusion handling ability to existing anchor-free detection network. OCSP integrates prior information into the network to handle occlusion. This high-fusion prior information gives the detector a hint about identifying the structural features of pedestrians. OCSP becomes more robust with the prior information which fusion the semantic head, the visible part, and the size and center body for each pedestrian. The detector enhances the perception of occlusion by predicting the semantic head and visible part. Besides, we design a head branch and add predicting the visible box to achieve a similar result. Experiments show that this fusion of prior information represents a suitable combination. We compare our OCSP with state-of-arts models on the Citypersons dataset, the proposed OCSP detector achieves the state-of-arts result on the CityPersons benchmark.
CITATION STYLE
Wang, H., Zhang, Y., Ke, H., Wei, N., & Xu, Z. (2021). Semantic Structural and Occlusive Feature Fusion for Pedestrian Detection. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 13031 LNAI, pp. 293–307). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-89188-6_22
Mendeley helps you to discover research relevant for your work.