摄像头网络中车辆检测和识别方法的研究

发布时间：2018-06-17 22:43

本文选题：车辆检测 + Faster　；参考：《大连海事大学》2017年硕士论文

【摘要】：近些年来,在视频监控联网、高清化技术的推动下,交通行业视频监控业务的数据量快速增长。利用计算机视觉技术处理交通视频得到有效的信息已经逐渐被重视,并且根据监控视频内容自动计算车辆运行轨迹已经可行。本文针对摄像头网络中车辆检测和识别问题进行了研究。对于摄像头网络,既需要对单摄像头下视频处理,又需要建立多个摄像头间的联系,从而完成在摄像头网络中对车辆进行连续的追踪。本文主要完成了以下工作:使用 Faster Region with Convolutional Neural Network feature 网络,在原有检测模型的基础上,重新标记数据集,对网络进行调优,重新训练车辆检测模型。在单摄像头下,提出了基于重叠面积率的车辆追踪算法,该算法利用相邻两帧视频帧中目标车辆的重叠面积率判定是否属于同一轨迹,对于车辆被遮挡以及Faster R-CNN检测失败或漏检的情况,加入卡尔曼滤波算法。利用卡尔曼滤波的预测机制,在进行重叠面积率计算后,若有未匹配成功的追踪器或车辆,则使用预测值进行匹配,这样可以尽量避免因检测失败或车辆被遮挡而导致的追踪车辆失败的情况。除此之外,为更直观的表达追踪结果,使用单应矩阵对追踪结果进行了可视化。在多摄像头下,针对不同摄像头下光照、拍摄角度等不同使得车辆再识别难度加大这一问题,本文根据摄像头间的时空关系、车型属性以及车辆Convolutional Neural Network特征建立了车辆再识别模型。其中利用摄像头的地理位置可以得到摄像头的空间信息,对视频统计可以得到车辆在摄像头间的转移时间概率密度分布,通过GoogLeNet网络重新训练车型检测模型,并结合VggNet网络提取到的车辆CNN特征,从而在不同摄像头下的监控视频中对目标车辆进行了再识别。
[Abstract]:In recent years, with the promotion of video surveillance network and high-definition technology, the traffic industry video surveillance business data volume is growing rapidly. Using computer vision technology to process traffic video to get effective information has been paid more and more attention, and it is feasible to automatically calculate the vehicle track according to the content of surveillance video. In this paper, the problem of vehicle detection and recognition in camera network is studied. For the webcam network, it is necessary to process the video under the single camera and to establish the connection between several cameras, so that the vehicle can be tracked continuously in the webcam network. The main work of this paper is as follows: using the Faster region with Convolutional Neural Network feature network, on the basis of the original detection model, the data set is re-marked, the network is optimized, and the vehicle detection model is retrained. A vehicle tracking algorithm based on overlapped area rate is proposed under a single camera. The algorithm uses the overlap area rate of the target vehicle in two adjacent frames to determine whether it belongs to the same track or not. In the case of vehicle occlusion and Faster R-CNN detection failure or miss detection, Kalman filter algorithm is added. The prediction mechanism of Kalman filter is used to calculate the overlap area rate. If there is a tracker or vehicle that has not been matched successfully, the prediction value is used to match. This can avoid tracking failure due to detection failure or vehicle occlusion. In addition, in order to express the tracking results more intuitively, the monoclinic matrix is used to visualize the tracking results. Under the multi-camera, aiming at the problem that different illumination and shooting angle under different cameras make it more difficult to recognize the vehicle again, according to the space-time relationship between the cameras, Vehicle rerecognition model is established by vehicle attributes and vehicle volume neural network features. The spatial information of the camera can be obtained by using the location of the camera, the probability density distribution of the transfer time between the cameras can be obtained by the video statistics, and the model of vehicle detection can be retrained through Google LeNet network. Combined with the features of vehicle CNN extracted from VggNet network, the target vehicles are rerecognized in the surveillance video under different cameras.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】