Aerial Video Multi-target Detection with Memory Module *

Abstract

Multi-target detection for aerial video is widely applied for both military and civilian. A multi-target detection algorithm on aerial video with memory module is proposed in this paper. This method uses memory module to connect time-scale information, so that the information of past frames can be used to improve the detection result of the current frame. ShuffleNet unit which is based on depthwise convolution and channel mixing was been used in the feature extraction network. The experimental results on VisDrone dataset show that the detection algorithm presented in this paper possesses fast detection speed and higher accuracy.