多特征融合目标跟踪算法研究

Research on Multi-Feature Fusion Target Tracking Algorithm

针对目标跟踪过程出现的遮挡、目标形变、复杂场景等挑战,提出了一种结合记忆机制的精确度较高的视觉目标跟踪算法。该算法首先对目标初始化,提取目标的HOG特征、纹理特征以及CN特征构建特征模型,弥补了单一特征可能造成跟踪不准确的问题。然后将模型和构建的滤波器进行相关滤波操作得到目标搜索区域的最大的响应位置,在后续的视频帧中采用包含了瞬时、短时、长时记忆空间的三层旋转圆记忆模型对目标模板进行更新,有效的增强了算法的跟踪能力。最后在OTB50、TC128、OTB100实验基准数据集上表明,与提到的KCF、DCF_SC、MOSSE_CA、SAMF_AT、STAPLE_CA算法进行对比,该算法在精度和准确度方面提高了,能够较准确的对目标进行跟踪,在目标跟踪领域具有重要的研究价值。

Aiming at the challenges of occlusion,target deformation,complex scenes and other challenges in the target tracking process,a high-precision visual target tracking algorithm combined with a memory mechanism is proposed.The algorithm first initializes the target and extracts the target’s HOG feature and texture features and CN features construct a feature model to make up for the inaccurate tracking problem caused by a single feature.Perform related filtering operations on the model and the constructed filter to obtain the maximum response position of the target search area.In the subsequent video frames,a three-layer rotating circle memory model containing instantaneous,short-term and long-term memory spaces is used to perform the target template.The update effectively enhances the tracking ability of the algorithm.The OTB50,TC128,OTB100 experimental benchmark data set shows that compared with other algorithms mentioned in the article,this algorithm can track the target more accurately,which has important research value.