Lightweight Real-time Object Detection Model for UAV Platform

Abstract

Real-time detecting objects on captured images on UAV (Unmanned Aerial Vehicle) platforms, rather than barely transmitting images back to supporting equipment for post-processing, is a core requirement for advanced UAV applications. However, due to limited computing capacity and memory of UAV platforms, it is very challenging to deploy real-time detection models on them. In addition, there are more small objects in aerial images, which makes it more difficult to detect accurately. To solve these problems, this paper brings dense connection to Yolo(You Only Look Once)v3 network, and proposes Yolo-LiteDense model. The backbone of Yolo-LiteDense is densely connected, which improves the performance of feature extraction. Then, we enforce channel pruning to Yolo-LiteDense model by pruning less informative channels with less scaling factors. After pruning, parameters and weight size of the model are compressed significantly, and inference time is also shortened. Evaluation results on VisDrone2018-DET show that parameters and weight size of Yolo-LiteDense are 83% and inference time is 30% less than Yolov3-SPP with comparable average precision. In addition, this paper also proposes the lighter version of Yolo-LiteDense, Yolo-DenseNano. Parameters and weight size of Yolo-LiteDense are 70% less than Yolov3-tiny with 2.68 times greater average precision.