基于光流法的车辆检测与跟踪

发布时间：2019-04-23 13:04

【摘要】：近年来,在智能交通领域中,基于视频的车辆检测跟踪已成为一个重要的研究方向。光流法在视频运动分析等领域具有极其重要的基础地位,能得到目标详细的二维运动信息,但光流算法计算的不适定及不能实时运算等问题限制了其广泛应用,因此研究出一种能实时执行且精度较高的光流算法对推广此算法在实时领域的应用显得尤为重要。若能将实时光流技术应用于智能交通领域,对智能交通的发展也大有裨益。本文详细阐述、分析了微分光流法的基本原理,并据此提出了多帧光流模型和其线性化求解方法,简化了光流的求解过程并提高了其估算精度;选择GPU平台做算法移植,解决光流的实时计算问题;根据所得实时稠密光流给出了车辆检测跟踪方案。首先,时域相关性在视频分析中具有重要作用,但在估算光流时这一特性却很少被应用。针对这一情况,提出在H-S光流模型基础上引入前向帧并加入时域相关性约束,从而构造出多帧光流模型。同时,针对能量泛函线性化求解过程异常复杂的情况,提出运用迭代重加权最小二乘法(IRLS)简化这一求解过程。其次,根据光流算法在像素间耦合性低这一特点,选择对其进行GPU平台移植。在计算统一设备(CUDA)架构下通过多线程并行执行方式同时对多个像素的光流值进行估算。在估算结果精度相当的情况下,在GPU上执行时间远小于CPU上执行时间。对于分辨率为640*480的视频图像可以达到实时性运算,能满足一般的应用要求。最后,给出一种改进的车辆检测跟踪方案。一方面,方案使用GPU平台计算所得实时稠密光流,相较于特征光流法和区域稠密光流法,可获得更加准确的全局运动信息,相较于帧差法等运动检测算法,也能获得更好的运动目标提取效果;另一方面,方案根据光流场所得到的速度矢量对车辆帧间的位置进行预测和匹配,能够对车辆在跟踪过程中常见的状态变化进行判断和处理,在实验中取得了预期的效果。
[Abstract]:In recent years, video-based vehicle detection and tracking has become an important research direction in the field of intelligent transportation. Optical flow method plays an important role in the field of video motion analysis. It can obtain the detailed two-dimensional motion information of the target. However, the problems such as the ill-posed calculation of optical flow algorithm and the impossibility of real-time operation limit its wide application. Therefore, it is very important to develop a real-time and high-precision optical flow algorithm to extend the application of this algorithm in real-time domain. If the real-time optical flow technology can be applied to the field of intelligent transportation, it will also be of great benefit to the development of intelligent transportation. In this paper, the basic principle of the differential optical flow method is analyzed in detail. Based on this, a multi-frame optical flow model and its linearization method are proposed, which simplifies the process of solving the optical flow and improves its estimation accuracy. The algorithm is transplanted on GPU platform to solve the problem of real-time calculation of optical flow, and the vehicle detection and tracking scheme is given according to the dense real-time optical flow obtained. First, time-domain correlation plays an important role in video analysis, but it is rarely used to estimate optical flow. In order to solve this problem, a multi-frame optical flow model is constructed by introducing forward frame and time-domain correlation constraint on the basis of HES optical flow model. At the same time, the iterative reweighted least square method (IRLS) is proposed to simplify the solution process in view of the complexity of the linearization process of the energy functional. Secondly, according to the low coupling between pixels, the optical flow algorithm is transplanted on GPU platform. The optical flow values of multiple pixels are estimated in parallel multi-thread execution mode under the (CUDA) architecture of computing unified device at the same time. Under the same precision, the execution time on GPU is much shorter than that on CPU. The video image with a resolution of 640 脳 480 can achieve real-time operation and can meet the general application requirements. Finally, an improved vehicle detection and tracking scheme is presented. On the one hand, the scheme uses the GPU platform to calculate the real-time dense optical flow. Compared with the characteristic optical flow method and the region dense optical flow method, the scheme can obtain more accurate global motion information, compared with the frame difference method and other motion detection algorithms. It can also get better effect of moving object extraction. On the other hand, the scheme predicts and matches the vehicle frame position according to the velocity vector obtained from the optical flow field, which can judge and deal with the common state changes in the tracking process of the vehicle, and achieves the expected effect in the experiment.
【学位授予单位】：江苏科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;TP391.41

【参考文献】