基于深度学习物体检测的视觉跟踪方法

A visual tracking method via object detection based on deep learning

提出了一种基于深度学习物体检测的视觉跟踪方法。该方法利用深度学习在特征表达上的优势,采用基于回归的深度检测模型SSD(Single Shot Multibox Detector)提取候选目标,并结合颜色直方图特征和HOG(Histogram of Oriented Gradient)特征进行目标筛选,实现目标跟踪。为了提升深度检测模型的物体检测性能,文中构建了多尺度目标搜索图,可在一张图上实现不同尺度的目标检测。在标准跟踪测试库上选取八个具有代表性的跟踪视频序列,并选取六种具有代表性的跟踪方法进行了对比测试。结果表明,文中所提方法在跟踪效果上,整体优于参与对比的其他算法,且对于物体姿态变化、尺寸变化、旋转变化、光照变化、复杂背景杂波等影响因素具有较好的鲁棒性。

A visual tracking method via object detection based on deep learning was proposed. In consideration of the advantages of deep learning in feature representation, deep model SSD（Single Shot Multibox Detector） was used as the candidate object extractor in the tracking model. Simultaneously, the color histogram feature and HOG（Histogram of Oriented Gradient） feature were combined to select the tracking object. In the process of tracking, multi-scale object searching map, which was applied to implement the object detection in different scales, was built to improve the detection performance of deep learning model. In the experiment of eight respective tracking video sequences in the baseline dataset,compared with six typical tracking methods, the proposed method has better performance in tracking effect, and has better robustness in the tracking challenging factors, such as deformation, scale variation,rotation variation, illumination variation, and background clutters.