Video object detection based on the spatial-temporal convolution feature memory model

Abstract

Video object detection technology can improve the battlefield object search capability of tank fire control system. However, complex battlefield environment and faster speeds of tank and objects bring great challenge to video object detection. A video object detection method is proposed in this paper for the tank fire control system. Given the rich spatial-temporal information in the video and the large position deviation of the target in the adjacent video frames, a spatial-temporal convolutional feature memory model consisted of spatial-temporal convolution feature alignment mechanism and convolution gated recurrent unit is proposed to transmit and fuse the information of adjacent frames. Moreover, the feature extraction network and the detection sub-network are improved by the deformable convolution networks to increase the detection accuracy of deformed objects. To evaluate the proposed method, a database named TFCS VID including 1396 videos labelled for seven types of typical objects in the battlefield was developed. Compared to several other video object detection methods, the proposed method achieved excellent detection results on TFCS VID and could better meet the actual application requirements of equipment.