目的 尺度突变是目标跟踪中一项极具挑战性的任务，短时间内目标的尺度发生突变会导致跟踪要素丢失，使得误差积累导致跟踪漂移，为了更好地解决这一问题，本文提出了一种先检测后跟踪的自适应尺度突变的跟踪算法（KCF_YOLO）。方法 该算法在跟踪的训练阶段使用相关滤波跟踪器实现快速跟踪，在检测阶段使用YOLO V3神经网络。并设计了自适应的模板更新策略，采用将检测到的物体与目标模板的颜色特征和图像指纹特征融合后的相似度进行对比的方法从而判断是否发生遮挡，据此决定是否在当前帧更新目标模板。结果 为证明本方法的有效性在OTB2015数据集具有尺度突变代表性的11个视频序列上进行试验，试验视频序列目标尺度变化大至9.2倍，小至0.1倍，结果表明本算法平均跟踪精度为0.955，平均跟踪速度36帧/秒；与经典尺度自适应跟踪算法进行比较，精度平均提高31.74%。结论 本文使用相关滤波和神经网络在目标跟踪过程中先检测后跟踪的思想，提高了算法对目标跟踪过程中尺度突变情况的适应能力，实验结果验证了加入检测策略对后续目标尺度发生突变导致跟踪漂移起到了很好的纠正作用，和自适应模板更新策略的有效性。
Objective Scale sudden change is a challenging task in object tracking, and it will lead to tracking drift in a short time. In order to solve this problem better, an adaptive scale sudden change tracking algorithm (KCF_YOLO) is proposed in this paper. Method The algorithm uses correlation filter tracker to achieve fast tracking in the training stage of tracking, and YOLO V3 neural network in the detection stage. An adaptive template updating strategy is designed. The similarity between the detected object and the target template is compared to determine whether occlusion occurs, and then whether the target template is updated in the current frame is determined. Results In order to prove the validity of this method, 11 video sequences with scale mutation representativeness in OTB2015 dataset were tested. The target scale change of the test video sequence was magnified to 9.2 times and reduced to 0.1 times. The results show that the average tracking accuracy of this algorithm is 0.955, and the average tracking speed is 36 frames/second; and it is adaptive to the classical scale. The tracking algorithm should be compared with each other, and the average accuracy can be improved by 31.74%. Conclusion In this paper, the idea of correlation filtering and neural network is used to detect before tracking in the process of target tracking, which improves the adaptability of the algorithm to scale mutation in the process of target tracking. The experimental results verify that the detection strategy has a good corrective effect on tracking drift caused by subsequent target scale mutation. And the effectiveness of adaptive template update strategy.