Rotated Top-View Fisheye Person Detection Based on Dynamic Attention

Addressing the issues of image distortion and difficulty in accurate target detection in overhead fisheye surveillance images, this paper proposes a real-time overhead fisheye image human target detection algorithm based on multi-attention fusion. Firstly, the dataset for object detection in aerial images (DOTA) class data processing method, used for remote sensing image target detection, is introduced into fisheye image processing. Then, based on the industrial-grade single-stage target detection algorithm RTMdet, the Dy-Head module is introduced to perform fusion attention processing on the output features of the original Path Aggregation Feature Pyramid Network (PAFPN). Finally, a shared detection head improves the detection accuracy while reducing the network parameters. Cross-comparison validation with different datasets show that the detection algorithm in this paper can achieve an accuracy of 90.6% on complex datasets, and the FPS can be stably maintained at 26.7, demonstrate superior human target detection capabilities on fisheye images compared to traditional target detection methods.