注意力叠加与时序特征融合的目标检测方法

提出一种基于注意力叠加与时序特征融合的目标检测方法.在端到端目标检测(DETR)网络的基础上,依据注意力机制特性,使用注意力权重叠加的方式提取目标物像素级标识,用于实例轨迹的划分.为使目标检测与轨迹跟踪协同作用,通过时序特征融合的方式融合之前轨迹跟踪信息,调整当前帧目标检测效果,从而充分利用视频载体提供的时间维度信息.在公开数据集上,对文中方法进行验证,结果表明:文中方法能有效识别被遮挡的目标物,具有较强鲁棒性.

Object Detection Method of Attention Superposition and Temporal Feature Fusion

An object detection method of attention superposition and temporal feature fusion is proposed. Based on the end-to-end object detection(DETR)network, attention weight superposition is used to extract the object pixel-level identification for the division of the instance trajectory according to the characteristics of the attention mechanism. In order to cooperate the object detection and trajectory tracking, the previous track tracking information is fused by the temporal feature fusion to adjust the effect of current frame object detection, so as to make full use of the temporal dimension information provided by the video carrier. The proposed method is experimentally tested on the public data set. The results show that the method in this paper can effectively detect the blocked object and has stronger robustness.