Occluded Pedestrian Detection Techniques by Deformable Attention-Guided Network (DAGN)

Although many deep-learning-based methods have achieved considerable detection performance for pedestrians with high visibility, their overall performances are still far from satisfactory, especially when heavily occluded instances included. In this research, we developed a novel pedestrian detector using deformable attention-guided network (DAGN). Considering that may be deformed occlusions or under diverse poses, designed convolution an attention module (DCAM) to sample non-rigid locations, and obtained the feature map by aggregating global context information. Furthermore, loss function was optimized get accurate bounding boxes, adopting complete-IoU regression, distance IoU-NMS used refine predicted boxes. Finally, preprocessing technique based on tone mapping applied cope low visibility cases due poor illumination. Extensive evaluations were conducted three popular traffic datasets. Our method could decrease log-average miss rate (MR−2) 12.44% 7.8%, respectively, heavy occlusion cases, compared published state-of-the-art results of Caltech dataset. Of CityPersons EuroCity Persons datasets, our proposed outperformed current best about 5% in MR−2 cases.