基于相关滤波和轨迹预测的视频卫星动态目标跟踪算法

为解决卫星视频中有遮挡或相似目标情况下目标跟踪的问题,提出了一种改进的相关滤波算法,在跟踪框架中加入了干扰判别模块和基于神经网络的轨迹预测模型。通过比较平均峰值相关能量指标值与自适应阈值来判别跟踪器是否受到干扰。双向长短期记忆网络以目标历史轨迹编码为输入来完成轨迹预测,结合相关滤波输出结果和网络预测结果来确定目标的位置。实验表明,所提出算法的精度提升了2.10%,在有遮挡或相似目标等情况下仍具有较好的跟踪性能。

Video Satellite Moving Object Tracking Algorithm Based on Correlation Filter and Trajectory Prediction

For video satellite moving object tracking under the conditions of occlusion or similar objects, an improved correlation filter algorithm is proposed. The correlation filter algorithm is combined with the interference discrimination module and trajectory prediction model based on the neural network. The values of the average peak-to-correlation energy and the adaptive threshold are compared to discriminate whether the tracker is interfered. The bi-directional long-short term memory network takes trajectory embeddings as the inputs for trajectory prediction. The object in the video is located by combining the output of the correlation filter and the network. Experiments indicate that the proposed algorithm improves the tracking precision and success rate by 2.73% and 5.70%, and it could achieve better tracking performance in satellite videos even with occlusion or similar objects.