Real-Time Object Detection in UAV Vision based on Neural Processing Units

With the reduction of Unmanned Aerial Vehicle (UAV) hardware cost and the development of deep learning algorithm, the real-time object detection algorithm applied in UAV vision has great advantages in many fields. However, due to the limited energy consumption and computing power of embedded devices used in the drones and the variable object scales and complex backgrounds in the UAV vision restrict the applications in object detection based on the drones. In this paper, we optimized the generation of anchor boxes, introduced a new module to increase the receptive field to improve the detection of small targets, and used adaptively spatial feature fusion in the feature pyramid to increase feature fusion of multi-scale features. At last we pruned the model to make it lighter and faster, and got the Average Precision (AP) of 89.7% for UAV car aerial images and the speed of 35.7 FPS by running on Neural Processing Units (NPUs), which proves the feasibility of the intelligent object detection algorithm's efficient processing in hardware resource limited environment.