结合帧差法的尺度自适应核相关滤波跟踪

Scale Adaptive Kernel Correlation Filter Tracking Combined with Frame Difference Method

为处理复杂应用场景下核相关滤波器跟踪效果不理想的问题,提出了一种结合帧差法的尺度自适应核相关滤波跟踪算法。在训练得到相关滤波器后,借助帧差法来处理下一帧图像,获得目标的预测位置,扩充算法的检测区域;然后通过尺度池构建多尺度待检测图像块集,通过相关滤波器来求得最大响应,估计出目标的最佳位置和最佳尺度;最后利用平均峰值相关能量(average peak-to correlation energy,APCE)作为跟踪置信度指标,引入高置信度更新机制,在目标被遮挡时,停止更新模型,防止误差被积累,提高正确率。在OTB100数据集上与若干视觉跟踪算法进行了对比实验,改进算法的成功率和距离精度均表现最优,比KCF算法高出21.7个百分点和12.0个百分点。该算法在目标快速运动、尺度变化、遮挡等复杂场景下,均具有较强的精确性和鲁棒性。

To deal with the problem of unsatisfactory tracking effect of kernel correlation filter under complex application scenarios,a scale adaptive kernel correlation filter tracking algorithm combined with frame difference method is proposed.After training the correlation filter,by means of the frame difference method,the next frame of image is processed,the predicted position of the target is obtained,and the detection area of the algorithm is expanded.Then,the multi-scale set of image blocks is built through the scale pool,and the maximum response is obtained through the correlation filter,and the optimal position and scale of the target are estimated.Finally,the average peak-to correlation energy(APCE)is used as the tracking confidence index,and a high confidence update mechanism is introduced.When the target is occluded,the updating model is stopped to prevent errors from being accumulated and improve accuracy.Compared with several visual tracking algorithms on the OTB100 database,experiments are performed on the improved algorithm,and the success rate and distance accuracy of the improved algorithm are the best,which are 21.7% and 12.0% higher than the KCF algorithm.The proposed algorithm has high accuracy and strong robustness in complex scenes such as rapid target movement,scale change,and occlusion.