改进的KCF算法在车辆跟踪中的应用

Application of improved KCF algorithm in vehicle tracking

摘要:

针对核相关滤波算法(KCF)在复杂道路场景下难以应对因车辆尺度变化,遮挡及旋转而不能继续跟踪的问题,提出了一种新的跟踪方法来更好地实现复杂道路场景下的车辆跟踪。该方法借鉴快速分类尺度空间跟踪器(fDDST),采用一维尺度相关滤波器进行尺度估计。同时融合Kalman滤波器形成预测-跟踪-校准的跟踪机制。该机制结合遮挡处理能够保证系统在目标被严重遮挡时跟踪的准确性。在模型更新方面,在目标被遮挡时,自适应的调节学习率参数,及时纠正模型偏移、特征丢失等问题。实验结果表明,在复杂道路场景下车辆旋转 、遮挡及尺度变化时,均能有效地跟踪目标车辆,且具有良好的鲁棒性。

Abstract:

To cope with the failure in continuously tracking in complex road scenes caused by vehicle scale changes, occlusion and rotation with the kernel correlation filtering algorithm (KCF), a new tracking method is proposed to better realize vehicle tracking under complex road scenes. Making reference to the fast discriminative spatial tracker(fDSST), this method makes scale estimations by adopting the one-dimensional scale correlation filter. Meanwhile, the Kalman filter is used to set up a prediction-tracking-calibration tracking mechanism. In the aid of an occlusion processing, it could keep a high accuracy of the system even the target is severely occluded. In terms of model updating, the learning rate parameter is adaptively adjusted, and problems like model offset and feature lose are solved in time when the target is occluded. The experimental results show that the proposed tracking method can effectively track the target vehicle when the vehicle rotates, occludes and scales in complex road scenes, thus has good robustness.