Pedestrian Detection in Crowded Environments Using a Hybrid Deep Learning Model with Occlusion-Aware Attention

Abstract

Pedestrian detection in crowded environments is an important computer vision problem, which has been broadly applied to intelligent transportation systems, autonomous vehicle, smart surveillance and public safety. Yet, accurate pedestrian detection in dense scenes is still difficult because of the significant occlusion, overlap, scale variation of objects and complex background. While the accuracy of detection has been greatly enhanced by deep learning methods, current models are not able to detect partly visible pedestrians, resulting in missed detections and false positives. A hybrid deep learning model that combines the benefits of both SSI and CNN by incorporating an Occlusion-Aware Attention (OA) mechanism for enhancing pedestrian detection in dense scenarios is proposed. The proposed framework combines a Swin Transformer backbone for feature extraction, a Feature Pyramid Network (FPN) for multi-scale feature learning, and a customized Occlusion-Aware Attention module to enhance the detection of partially occluded pedestrians. To detect overlapping pedestrians efficiently, a YOLO based detection head is used and Soft Non-Maximum Suppression (Soft-NMS) is adopted to refine the detected overlapping pedestrians. The framework will be tested with the benchmark datasets: CrowdHuman, CityPersons and WiderPerson, and the metrics used will be: Precision, Recall, F1-Score, mAP@0.5 and inference speed. The proposed model is expected to achieve high accuracy in detection, low probability of missed detection in high occlusion rate and efficient solution for real-time pedestrian detection in complex urban environment.