深度残差特征与熵能量优化运动目标跟踪算法

针对模型更新的运动目标跟踪算法准确率、实时性和鲁棒性较低的问题,提出一种基于深度残差特征与熵能量优化的运动目标跟踪算法。通过深度残差网络从视频序列中提取深度残差特征,计算深度残差特征的熵能量,并通过二维核变换计算深度频率。由微分方程从深度频率中计算出深度平衡,通过极大似然估计出目标位置和速度等状态信息,完成对运动目标的跟踪。为了验证算法的可行性与有效性,在目标跟踪基准数据集(object tracking basis, OTB)上进行算法对比试验,验证各个算法在运动目标跟踪上的准确性和鲁棒性。试验结果表明,该研究提出的算法比当前最佳算法在运动目标跟踪的速度和位置准确性上都有显著的提升,通过深度残差特征的熵能量优化,使运动目标跟踪算法具有更好的灵活性和鲁棒性。

Object tracking algorithm based on deep residual features and entropy energy optimization

To solve the low rate of accuracy, real-time and robustness of object tracking algorithm based on model updating, a new algorithm based on deep residual features and entropy energy optimization was proposed. Deep residual features were first extracted from original video sequence by deep residual network. The entropy energy from deep residual features were calculated, and the deep frequency from entropy energy by two-dimension kernel transformation could be calculated, after that we got the deep balance by deep frequency with differential equation, and then the object state by MLE was estimated, including object position and speed. To validate the feasibility and efficiency of the proposed algorithm, the comparing experiments on the object tracking basis (OTB) dataset for the state-of-the-art algorithms were done, and the comparison results showed that the proposed algorithm had significant improvement on tracking accuracy and robustness. By using entropy energy optimization for deep residual features, the proposed algorithm had more flexibility and robustness for object tracking.