基于自适应模型的时空上下文跟踪

Visual Tracking Using Adaptive Structure Model Based on Spatio-Temporal Context

针对当前时空上下文目标跟踪算法存在的易发生模型漂移的问题,提出基于自适应模型的时空上下文跟踪算法。该算法通过对常规模板保存多个历史快照模型作为多模板,当历史快照模板估计到比常规模板适应性更强的结果时,立即对常规模板进行回滚,可有效提升时空上下文跟踪算法的鲁棒性,在目标快速运动、快速旋转、运动模糊和严重遮挡的情况下依然跟踪准确。在Tracker Benchmark v1.0测试集上与原时空上下文目标跟踪算法的对比实验表明,平均正确率由38.61%提高到42.02%,并将平均中心坐标误差从85.57降低到62.78,而平均帧速则从45.89 fps下降到36.64 fps,依然满足实时跟踪的要求,表明该算法在面对多种因素干扰的场景下,仍能完成稳定的实时跟踪。

Object tracking is one of the basic problems in the field of computer vision. There are many algorithms presented, and STC is a quite novel one. But the STC tracking method can＇t deal with model drift problem. To overcome this weakness, proposes an algorithm using adaptive structure model based on STC. This algorithm takes a set number of snapshots of normal template as snapshot templates, and saves them to snapshot set. When one of the snapshot templates gets an enough better outcome than normal template, the algorithm uses the snapshot template to roll back the normal template, which can effectively enhance the tracking robustness and keep accurate even when object suffers all kinds of interferences such as fast motion, in-plane rotation, motion blur, severe occlusion and so on. The experi-ment on Tracker Benchmark v1.0 dataset compared to STC shows that the average accuracy is improved from 38.61% to 42.02%, the average center location error is decreased from 85.57 to 62.78 and the average frame rate decreased from 45.89 fps to 36.64 fps. It still meets the requirements of real-time tracking. In conclusion, this algorithm can accomplish the real-time tracking steadily even when dis-turbed by all kinds of interferences.