复杂交通视频场景中的车辆轨迹提取及行为分析

发布时间：2017-12-28 07:00

本文关键词：复杂交通视频场景中的车辆轨迹提取及行为分析　出处：《长安大学》2016年博士论文　论文类型：学位论文

【摘要】：基于视频的车辆运动轨迹提取及行为分析作为一个多学科交叉融合形成的研究领域,涵盖了数字图像处理技术、人工智能以及模式识别等多学科知识。然而,由于该领域研究对象复杂,涉及学科众多,目前仍有很多难点问题亟待解决。复杂交通场景中运动车辆检测、跟踪和行为识别一直是该领域研究的热点和难点,许多方法和技术还不够成熟和完善。本文围绕基于视频的车辆轨迹提取与行为分析中的目标车辆检测、运动车辆跟踪、车辆轨迹相似性度量和轨迹聚类等关键问题进行了深入研究,取得了以下主要研究成果:1)针对复杂交通场景下的车辆目标检测,本文提出一种基于车辆对称特征和阴影特征的车辆目标检测方法。该方法在SURF特征提取算法的基础上,利用水平镜像矩阵构造新的SURF特征描述算子。由于视觉上具有对称特性的特征点处于不同尺度时,其匹配误差会比较大。因此,本文从减少Haar特征累加次数和降低尺度对特征点表示的影响两方面入手,对S-SURF算法进行改进和优化。然后采用优化的S-SURF算法对车辆的对称特征进行提取,并利用车辆对称特性对车辆的中心位置进行定位,最后,根据车辆底部阴影特征对车辆目标进行识别和区域定位。实验结果表明,该方法利用局部不变特征集合来描述车辆目标,有效地避免了复杂场景下的目标分割难题,同时简化了局部特征检测方法中的聚类问题,复杂度较低,且具有较高的准确性。2)运动车辆的可靠稳定跟踪是车辆轨迹提取的关键。本文提出一种融合特征匹配和光流法的车辆目标跟踪方法,该方法在基于双向可逆性约束的KLT算法的基础上,构造新的偏移量估算方法,对稳定性较差的特征点进行剔除,提高了特征点跟踪的可靠性和稳定性。同时,采用SURF特征匹配算法作为补偿机制对目标特征点集进行更新和校正。最后,结合初始帧中特征点之间相对位置和相对角度的关系,确定当前帧中目标的尺度变化和旋转变化,并采用层次聚类的方法,对特征点进行聚类,以此删除异常特征点,从而确定当前帧中的目标区域。该算法将两个匹配策略相结合,既提高了跟踪算法的稳定性,也很好地解决了目标在被跟踪过程中发生的形变、部分遮挡等问题,对目标的尺度和旋转变化也具有较强的鲁棒性。3)运动轨迹的相似性度量是轨迹聚类过程中的一个核心问题,由于车辆轨迹的复杂性和多样性,现有度量方法都有其局限性。本文提出一种融合多特征和编辑距离的轨迹相似性度量方法。该方法在EDR编辑距离的基础上,结合轨迹点的速度和方向特征,对轨迹进行分段处理,并给具有不同特征意义的分段赋予不同的编辑操作代价值。最后,对基于分段表示的IEDR算法进行了进一步的定义和分析。该算法保留了EDR算法的允许时间伸缩、抗噪性等优点的同时,将轨迹点的位置、速度和方向特征合理地融入到车辆运动轨迹的相似性度量中,进一步提高了轨迹相似性度量的准确性和鲁棒性。4)车辆行为模式学习的目的是提取出具体交通场景的常态运动模式,从而为车辆异常行为识别研究提供前提条件。本文提出一个基于增量式DPMM的贝叶斯最大后验概率估计方法的轨迹聚类模型。该方法采用DFT系数作为轨迹的特征表示方法,提出一种基于DPMM的轨迹聚类方法,并在此基础上,对Gibbs抽样过程进行改进,以已分类轨迹作为先验知识,对新增轨迹类别进行划分。同时,在分类过程中,学习轨迹的常态运动模式,通过运动模式和方向模式匹配策略,对车辆异常行为进行判别。该算法不需要训练样本,而且随着新增轨迹的到来而变化,聚类模型能够实现自适应变化及模型参数学习和分类数目自动更新的任务,很好地解决了由于交通异常行为的不可预知、不常发生性引起的数据稀疏情况下的模型训练困难问题。同时,利用已有聚类结果,将每次新增轨迹划分到已有类别或新类中,不需要每次对所有轨迹进行重新聚类,聚类效率大大提高。
[Abstract]:Video based vehicle motion trajectory extraction and behavior analysis is a research field formed by multidisciplinary fusion, which covers multidisciplinary knowledge such as digital image processing technology, AI and pattern recognition. However, because of the complexity of the research object in this field and many subjects, there are still many difficult problems to be solved. Moving vehicle detection, tracking and behavior recognition in complex traffic scenes has been a hot and difficult topic in the field. Many methods and technologies are not mature enough. This paper focuses on the analysis of the target vehicle vehicle trajectory extraction and behavior in video detection, vehicle tracking, vehicle trajectory similarity measure and the key problem of trajectory clustering based on in-depth research, the main results are as follows: 1) for vehicle target detection under complex traffic scene, this paper proposes a vehicle detection method for target the vehicle features and shadow features based on symmetry. On the basis of the SURF feature extraction algorithm, this method constructs a new SURF feature description operator by using the horizontal mirror matrix. Because the feature points of the visual symmetry are in different scales, the matching error will be larger. Therefore, this paper improves and optimizes the S-SURF algorithm from two aspects: reducing the number of Haar feature accumulating times and reducing the impact of the scale on the representation of the feature points. Then, the optimized S-SURF algorithm is used to extract the symmetrical characteristics of the vehicle, and the vehicle's central location is located by the symmetry characteristics of the vehicle. Finally, the vehicle's target is identified and located according to the shadow feature of the vehicle bottom. The experimental results show that the proposed method uses local invariant feature set to describe vehicle targets, effectively avoids the problem of target segmentation in complex scenes, and simplifies the clustering problem in local feature detection, with low complexity and high accuracy. 2) the reliable and stable tracking of the moving vehicles is the key to the vehicle trajectory extraction. The vehicle target tracking method this paper proposes a fusion feature matching and optical flow method, the method based on KLT algorithm of bidirectional reversible constraints on the structure of the new method to estimate the offset, poor stability of feature points are removed, improves the feature tracking reliability and stability. At the same time, the SURF feature matching algorithm is used as compensation mechanism to update and correct the target set of feature points. Finally, combined with the relationship between the feature points in the initial frame relative position and angle of the scale change to determine the target in the current frame and rotation, and the method of hierarchical clustering, clustering of feature points, in order to remove abnormal points, so as to determine the area of the object in the current frame. The algorithm combines two matching strategies, which not only improves the stability of tracking algorithm, but also solves the problems of deformation and partial occlusion during target tracking. It also has strong robustness to the scale and rotation of targets. 3) the similarity measurement of trajectory is a core problem in trajectory clustering. Due to the complexity and diversity of vehicle trajectories, existing metric methods have their limitations. In this paper, a method of trajectory similarity measurement is proposed, which combines multiple features and edit distance. Based on the edit distance of EDR, combined with the speed and direction characteristics of track points, the method segmented the trajectories and assigned different editing operations to different feature segments. Finally, the IEDR algorithm based on piecewise representation is further defined and analyzed. The algorithm preserves the advantages of EDR algorithm such as the allowed time expansion and noise immunity. Meanwhile, it integrates the location, velocity and direction characteristics of trajectories reasonably into the similarity measurement of vehicle trajectories, which further improves the accuracy and robustness of trajectory similarity measurement. 4) the purpose of vehicle behavior model learning is to extract the normal motion pattern of specific traffic scene, thus providing the precondition for the research of vehicle abnormal behavior recognition. This paper presents a trajectory clustering model based on an incremental DPMM based Bayesian maximum a posteriori probability estimation method. In this method, the DFT coefficient is used as the characteristic expression method of trajectory. A trajectory clustering method based on DPMM is proposed. Based on that, the Gibbs sampling process is improved, and the classified track is used as a priori knowledge to classify the new trajectory categories. At the same time, in the classification process, the normal motion model of learning trajectory is used to discriminate the abnormal behavior of the vehicle through the motion pattern and the direction pattern matching strategy. The algorithm does not need training samples, and changes with the arrival of new track, clustering model can realize adaptive and parameter learning and classification number of automatic update tasks, is a good solution to the traffic abnormal behavior is unpredictable, infrequent training difficult model data sparseness problem caused by the situation. At the same time, we use existing clustering results to divide every new trajectory into existing categories or new classes, and do not need to re track all trajectories at any time, so the clustering efficiency is greatly improved.
【学位授予单位】：长安大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】