基于深度学习的目标跟踪方法及其实现

Object Tracking Method Based on Deep Learning and Its Implementation

提出一种基于深度学习的目标跟踪方法。首先采用一种直接预测边界框坐标和类别置信度的SSD(Single Shot Multibox Detector)算法提取候选目标,然后结合颜色直方图和方向梯度直方图特征计算待跟踪目标与候选目标的巴氏系数,取最接近者作为匹配目标,最后基于卡尔曼滤波剔除误跟踪目标。实验表明该文所提方法是可行的,具有准确性高、实时性能好等特点。

This paper proposes an object tracking method based on deep learning.Firstly,the potential candidates are extracted by SSD (Single Shot Multibox Detector) algorithm which directly predicts the bounding box coordinates and the class confidence.Then,the Bhattacharyya coefficients of the object to be tracked and the candidate objects are calculated by combining the color histogram with the Histograms of Oriented Gradients (HOG) features.The candidate with maximum similarity is chosen as the best matching object.Furthermore,the outliers are eliminated leveraging Kalman filtering.