Onboard Real-Time Object Detection for UAV with Embedded NPU

Unmanned aerial vehicles (UAVs) endowed with the function of computer vision are widely applied. There is an urgent need to implement real-Time object detection. However, the limited memory and computing capacity of current embedded devices prevent the deployment of deep learning on UAVs. Therefore, this paper proposes an efficient onboard object detection system based on an embedded NPU device. It can be deployed on a UAV to perform real-Time scenario analysis. Firstly, a deep-learning network structure based on YOLOV3-Tiny is designed for object detection. Then it is compressed by a novel pruning strategy to obtain a 'slim' object detector. Finally, a detection-reinitialization mechanism is introduced to realize the robust output of detection. The experimental results demonstrate that our model and algorithm achieve efficient performance of detection compared to the unoptimized benchmark. The proposed model achieves up to 0.591 mAP with a 6M parameter size. The complete system is tested in both simulation and real environments of the aerial manipulator grasping task. The effective detection frame ratio reaches 90%, and FPS achieves about 25.