提出了一种基于深度学习检测算法YOLOv3的目标跟踪算法。该算法利用深度学习模型在目标特征提取上的优势, 采用基于回归的YOLOv3检测模型提取候选目标, 同时结合目标颜色直方特征和局部二值模式直方特征进行目标筛选, 实现对目标的跟踪。为了提高算法的性能, 还提出了一种K邻域搜索方法, 可以针对选定目标进行邻域检测。实验结果表明, 提出的目标跟踪算法跟踪效果很好, 综合表现比4种对比算法提高了80%左右, 同时在目标物体光照变化、姿态变化、尺寸变化、旋转变化等复杂情况下有很好的鲁棒性。

An object tracking algorithm is proposed based on the deep learning detection algorithm of YOLOv3 (YOLOv3:An Incremental Improvement), which utilizes the advantages of deep learning model in target feature extraction, and extracts candidate targets by using regression-based YOLOv3 detection model. The target color histogram feature and Local Binary Pattern (LBP) feature are also used for target screening, thus to implement object tracking.At the same timea method called K-neighbor searching is presented to improve algorithm performance, which performs neighborhood detection for the selected targets. Experimental results show that the proposed algorithm has a good tracking performance, with an overall performance improved by about 80% in comparison with the four contrast algorithms, and has good robustness in the complex situations of illumination changing, posture changing, size changing and rotation of target object.