Research on the application of artificial intelligence and multi-scale image fusion technology to pedestrian detection in complex street view

With the increasing face imaging data and the advancement of artificial intelligence (AI) technology, computer-aided monitoring systems are crucial for pedestrian detection in dense street view. However, due to occlusion and small pedestrian scale, pedestrian false alarms and missed detection problems become more and more serious. Therefore, this paper proposes a pedestrian detection model, YOLOv10s-pedestrian. Firstly, CA attention is introduced to redesign the MBConv module, resulting in an efficient MB-CANet backbone for pedestrian feature extraction, enhancing the accurate localization of densely occluded pedestrians. Secondly, a novel C2FN structure was created to reduce the number of parameters while improving the model's accuracy. Additionally, inspired by the BiFPN feature fusion concept, a Bi-C2FN-FPN network structure is proposed to effectively fuse features from different depth sources, strengthening feature fusion and improving pedestrian detection accuracy. Finally, the MPDIOU loss function replaces the original CIoU loss function to enhance anchor box localization. Experimental results demonstrate that the proposed model achieves a mAP50 of 95.6% on the WiderPerson pedestrian detection dataset, which is a 6.1% improvement over the original model, with a recall rate of 86.2%, showcasing excellent detection performance. Compared to several mainstream object detection models, the proposed model also exhibits superior performance.