基于二进制光流描述子的运动目标提议

MOTION OBJECT PROPOSAL BASED ON BINARY OPTICAL FLOW DESCRIPTOR

针对视频目标检测提议框存在大量冗余的问题,提出二进制光流描述子,实现视频连通目标的分离提议。通过分析目标局部运动一致性,设计二进制光流描述子,构建目标提议模型,然后利用线性支持向量机分类器求解运动目标提议参数。为充分利用二进制与或运算速度快的优势,采用二进制近似表示运动目标提议参数,快速高效地完成目标提议任务。实验针对Caltech行人数据库,快速地生成了少量高质量的运动目标提议窗口。实验结果优于现有目标提议方法。

In order to reduce the redundance of proposal windows in video object detection, a novel binary optical flow descriptor is proposed to achieve separate proposals of objects, we design the binary optical flow descriptor, for connected objects. By analysing the local motion coherence construct the object proposal model and then use linear SVM classifier to solve motion object proposal parameters. To make full use of the speed advantage of binary and/or calculation, the motion object proposal parameters are represented approximately in binary to complete object proposal tasks quickly and efficiently. The experiment on the Cahech Pedestrians dataset is carried out, and a small number of high-quality motion object p is superior to current object roposal windows are generated quickly. Experimental results show that the proposed method proposal methods.