【作者】 王昕；

【导师】 王常虹；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2016， 硕士

【摘要】 目标检测与跟踪技术是计算机视觉、视频图像监控与模式识别领域中的核心课题之一,主要的目标为获取跟踪目标的位置信息与运动参数(如速度、加速度等)并得出目标在图像序列中的运动轨迹,为进一步的图像理解、目标行为分析等做好基础,从而完成更高级的课题任务。本文以道路上的车辆为研究对象,针对不同场景下的车辆跟踪问题进行了研究。论文的主要研究是基于跟踪-学习-检测的思想,通过检测器来增强跟踪器的稳定性,并使用跟踪器的结果作为正负样本在线学习更新检测器参数。论文主要包含三个部分:迭代跟踪器算法研究、检测器研究与在线学习方法研究。首先,根据跟踪实际需求,设计跟踪器模块。文中首先比较了三种不同的算法,并在其基础上分析改进,提出了一种光流Mean-shift融合的方法来对目标进行跟踪,并与一般光流法比较,给出实验分析,得出较好的实验结果。其次,设计目标检测模块。分别尝试了Haar特征分类器和级联分类器检测目标,并在级联分类器中改进滑动窗口搜索机制,并应用ViBe前景提取算法作为第二步弱分类器,采用随机蕨类分类器作为第三级分类器,最后使用模板匹配得出检测目标位置。在此基础上,对算法中的学习融合模块做出设计。基于运动连续性原则,以跟踪结果邻近滑动窗口图像块作为正样本,其余为负样本,对检测器参数进行了在线学习,得到级联检测器的模板匹配方法中的样本类更新以及最后,实验使用PETS2000数据库和实际交通监控视频验证算法。实验结果表明在多个场景下,达到了一定的准确率,但是跟踪速度不能满足实时性要求;与此同时,在跟踪目标产生形变、旋转、尺度变化以及遮挡时可以稳定跟踪目标。更多还原

【Abstract】 Target detection and tracking technology is one of the core issues in the field of computer vision, video surveillance and pattern recognition. The main goal for achieving tracking target position information and movement parameters(such as speed, acceleration, etc.) and draw the target trajectory in image sequence, so as to get prepared for image understanding, target behavior analysis which are more advanced tasks in video surveillance. This paper takes the vehicle on the road as research object, and focus on the research on vehicle tracking in different scenarios. The main research is based on the idea of tracking by detection that the detector enhances the robustness of tracker and the result of tracker is used as the samples to train detector online. The thesis mainly consists of three parts: the study on recursive tracking algorithm, the detector and the online learning method.First, we design the tracker module. This paper compared three different algorithms, and after analysis and improvement, a fusion of optical flow and Mean-shift method is proposed to track targets. The classic optical flow method is compared in experiments and our method is proved to get a better result.Secondly, a target detection module is designed. We, respectively, try the Haar-like feature classifier and cascade classifier to detect target. In the cascade classifier, we improved sliding window search mechanism, and applied Vibe foreground extraction algorithm as the second stage in cascade classifiers. The third stages of cascade classifier is implemented by the random fern method. In the last stage of the cascade classifier, we use template matching method to achieve the detector location.Then, the online learning module is designed. Based on the principle of motion continuity, we regard the sliding windows adjacent to tracking result as positive samples, and the rest as negative ones, and then the detector parameters are learned online.Finally, our experiments use PETS2000 database and the actual traffic monitoring video to verify the proposed algorithm. The results show that in a number of different scenarios, the method achieved a certain accuracy, but the tracking speed cannot meet the real-time requirements. Meanwhile, the proposed algorithm performs well in the situation that rotation, occlusion, deformation and scale changing occurs.更多还原