Autonomous Aerial Vehicle Object Detection Based on Spatial Perception and Multiscale Semantic and Detail Feature Fusion

Owing to changes in the spatial position of Autonomous Aerial Vehicle (AAV) aerial images and limited platform resources, most existing AAV aerial image detection models have low accuracy, and it is difficult to achieve a good balance between detection performance and lightweight. To solve the above problems, an object detection model of AAV aerial images based on YOLOv8s, called BSDS-YOLOv8s, is proposed. In the proposed model the Dynamic Head (DyHead) is used to replace the detection head of YOLOv8s firstly, which can improve the spatial perception ability of the detection head of the model. Second, to improve the detection performance of DyHead and make the model lightweight, a new feature pyramid network (SDI-MBiFPN) in the Neck of YOLOv8s is proposed, which contains a Multiscale Bidirectional Feature Pyramid Network (BiFPN), called MBiFPN, and a feature fusion method based on the redesigned Semantic and Detail Fusion (SDI). Finally, a method that integrates Soft Non-Maximum Suppression (Soft-NMS) with Generalized Intersection over Union (GIoU), called GIoU-Soft-NMS, is proposed to enhance the model’s post-processing capability and reduce the missed and false detection rates, thereby further improving the detection accuracy of the model. The experimental results showed that the mAP0.5 and mAP0.5:0.95 of BSDS-YOLOv8s reached 49.6% and 34.1%, respectively, on the VisDrone2019 dataset. This is an improvement of 9.2 and 9.8 percentage points over YOLOv8s, respectively. The number of parameters and the floating point operations (FLOPs) were 8.03M and 26.6G, 27.8% and 7.6% less than those of YOLOv8s, respectively.