Bi-Level Routing Attention–Feature Fusion Net for Oriented Small Object Detection using UAV Infrared Imaging in Traffic Monitoring

With the rapid development of UAV (Unmanned Aerial Vehicle) and infrared imaging technology, object detection based on the resultant images has demonstrated its potential application value in various fields; however, the existing oriented road object detection methods do not meet practical requirements in terms of detection accuracy, multi-scale object detection, etc., and there is still space for improving existing algorithms. Therefore, we propose bi-level routing attention–feature fusion net (BRAFF net) for oriented small object detection with UAV infrared images in traffic monitoring. We validate the effectiveness of the proposed algorithm using the HIT-UAV and DroneVehicle datasets. The experimental results show that, with small differences in model size and parameter count, the proposed algorithm achieves a target detection accuracy of 87.9%, which is higher than those of existing mainstream oriented object detection methods, and meets the re-al-time requirements for road object detection. Additionally, the proposed method can be useful not only in the fields of road object detection and traffic monitoring but also has some referential significance in other object detection fields.