Research On YOLOv8 UAV Ground Target Detection Based On RK3588

In order to realize the deployment of object detection on edge computing devices more efficiently and quickly, and improve the ability of UAV to detect ground targets, this paper selects the Rockchip RK3588 chip to compile and quantitatively deploy the YOLOv8 object detection algorithm model, and proposes a multi-threaded resource interlock design through airborne computing inference optimization, so as to realize the convolutional neural network to improve the ability of UAV ground target detection. In terms of RK3588 inference speed analysis, after the optimization and improvement of multi-threaded optimization processing, when the number of threads is 3, the fps has basically reached about 100, and a method of knowledge distillation applied to the lightweight processing of the object detection model is proposed to quantify the original FP32 model. The number of detection targets and the detection confidence information of the two are counted, which shows that the detection accuracy and detection confidence of INT8 are basically unchanged.