A State-of-the-art High-Resolution Network based Dipolar Detector for Occluded Pedestrian Detection

Abstract

Applications including robotics, autonomous driving, assisted living, and surveillance all depend extensively on pedestrian detection. Despite the fact that several computer vision researchers have attempted to tackle the pedestrian detection problem, it remains unresolved. Scale, position, occlusion, lighting, and many other similar elements have an impact on how well the procedures work. So a novel dipolar modelling method based on the High Resolution Network (HRNet) has been proposed for an efficient pedestrian detector that resolves the occlusion issue. The two centre prediction branches built for the pedestrian detection head specifically changed the training data loss or ground truth. In order to improve the detection performance,the end-to-end optimise network's is designed as detection head and make use of the supportive outputs for the two focus projection branches. The network is compeled to focus more on the occurrences of occluded pedestrians by the proposed dipolar prediction technique. The experimental findings on the KITTI benchmark support the proposed dipolar detector's effectiveness in controlling occlusion in pedestrian identification