Real-Time object detector based MobileNetV3 for UAV applications

With the continuous progress of UAV (unmanned aerial vehicle) flight technology, more and more outdoor vision tasks begin to rely on UAV to complete, many of which require computer vision algorithms to analyze the information captured by the camera. However, it is difficult to deploy detectors on embedded devices due to the challenges among energy consumption, accuracy, and speed. In this paper, we propose an end-to-end object detection model running on a UAV platform that is suitable for real-time applications. Through the research of shufflenetv2 and mobilenetv3, a new feature extraction network structure is proposed. In order to improve the detection accuracy without losing the detection efficiency, a multi-scale fusion module based on deconvolution is added. Experiments show when deployed on our onboard Nvidia Jetson TX2 for testing and inference, our model combined with a modified focal loss function, produced a desirable performance of 21.7% mAP for object detection with an inference time of 17 fps.