融合多尺度和深度特征的快速运动视觉目标跟踪

Fast Moving Vision Target Tracking Based on Multi-scale and Depth Features

为了提高快速运动视觉目标跟踪的精度、效率和鲁棒性,提出融合多尺度和深度特征的快速运动视觉目标跟踪方法。通过分类器训练、候选区域检测以及模型更新三个阶段,完成核相关滤波跟踪算法。为适应视觉目标的多尺度变化,通过双线性插值函数将采集的子图转化为多尺度大小;提取子图特征,对多尺度核相关滤波器进行训练,完成快速运动视觉目标跟踪;利用堆叠多层自编码器构建深层神经网络,通过贪婪算法对输入数据重构,得到深度特征;将深度特征和方向梯度直方图HOG特征进行匹配融合,实现快速运动视觉目标跟踪。实验结果表明,该算法能够有效跟踪被遮挡的快速运动视觉目标,其跟踪准确率和跟踪成功率分别高达97.6%和98.2%,跟踪时间仅为4.5 ms,可以有效提高快速运动视觉目标跟踪精度和效率,增强鲁棒性。

In order to improve the accuracy,efficiency and robustness of fast moving visual object tracking,a fast moving visual object tracking method combining multi-scale and depth features was proposed.Through three stages of classifier training,candidate region detection and model updating,the kernel correlation filter tracking algorithm was completed.In order to adapt to the multi-scale changes of visual objects,the collected sub-images were converted into multi-scale sizes by bilinear interpolation function.Extract the sub-image features,train the multi-scale kernel correlation filter,and complete the fast moving visual target tracking.A deep neural network was constructed by using stacked multilayer self-coder.The input data was reconstructed by greedy algorithm to obtain depth features.The depth feature and directional gradient histogram HOG feature were matched and fused to achieve fast moving visual target tracking.The experimental results showed that the proposed algorithm could effectively track the occluded fast moving visual object,and its tracking accuracy and tracking success rate were as high as 97.6%and 98.2%respectively,and the tracking time was only 4.5 ms.It could effectively improve the tracking accuracy and efficiency of fast moving visual object,and enhance robustness.