交通场景中的车辆跟踪算法研究

发布时间：2018-09-17 14:31

【摘要】：智能化的交通解决方案已经成为缓解交通压力的有效手段,人工智能和大数据的蓬勃发展,给智能交通领域带来了新的活力和发展。与此同时,基于交通监控的智能化算法是当前交通信息采集的前沿发展方向和热点问题,而基于视频的车辆检测和跟踪任务则是交通信息采集的基础和关键工作,是多样交通参数获取的基础数据支撑。由于交通场景的复杂性,造成了车辆检测和跟踪中存在诸多挑战,例如遮挡、光照变化以及实时性要求等问题。近年来,深度学习技术在计算机视觉等领域取得了突破性进展,也使得通过视频分析方法实现交通状态分析和理解成为了可能。随着交通监控设备的大量应用,具有丰富交通信息的视频数据井喷式增长,因此快速、准确地车辆检测和跟踪方法对于交通信息获取,交通运行管理具有重要意义。本文基于实际道路中多维度的交通视频监控场景,以目前优秀的目标检测和跟踪算法为基础,对实现准确高效的车辆跟踪算法进行了深入研究,提出了多模块融合的车辆跟踪框架。本文主要完成了以下工作:首先,提出了基于深度神经网络的车辆检测方法。依据车辆及其周围环境信息,以候选区域提出网络和检测网络组成的网络结构为基础,利用深度学习框架Caffe,训练得到了鲁棒的车辆检测器。实验验证了算法在多种天气和交通场景中取得了良好的效果。其次,针对核化相关滤波跟踪方法在模型更新方面的缺陷,提出一种基于增量学习的模型更新方法。通过建立早期跟踪模型的快照集合,结合近邻帧收集的模型,采用增量更新的方式,构建了包含早期以及当前目标信息的主成分的跟踪模型。通过与多种模型更新方法对比,证明了增量更新的鲁棒性和适应性。最后,在车辆检测和相关滤波跟踪算法的基础上,提出了车辆跟踪算法。车辆检测器用于初始化多尺度的相关滤波跟踪算法和纠正跟踪失败。相关滤波跟踪算法用于实现短时间单车辆跟踪。通过基于轨迹片段置信度的关联方法,实现了检测和跟踪模块的有机融合,形成了更加完整的车辆轨迹,实现了复杂交通环境下的车辆跟踪。实验证明该跟踪框架实现了车辆速度和准确性的平衡,在实际道路监控视频中取得了优秀的效果。
[Abstract]:Intelligent traffic solution has become an effective means to relieve traffic pressure. The vigorous development of artificial intelligence and big data has brought new vitality and development to the intelligent transportation field. At the same time, the intelligent algorithm based on traffic monitoring is the forward development direction and hot issue of traffic information collection, and the task of vehicle detection and tracking based on video is the basis and key work of traffic information collection. It is the basic data support for obtaining various traffic parameters. Because of the complexity of traffic scene, there are many challenges in vehicle detection and tracking, such as occlusion, illumination change and real-time requirements. In recent years, deep learning technology has made a breakthrough in the field of computer vision, which makes it possible to realize traffic state analysis and understanding by means of video analysis. With the extensive application of traffic monitoring equipment, the video data with abundant traffic information is increasing rapidly and accurately, so it is very important for traffic information acquisition and traffic operation management to detect and track vehicles quickly and accurately. Based on the multi-dimensional traffic video surveillance scene in the actual road, based on the current excellent target detection and tracking algorithm, this paper makes a deep research on the realization of accurate and efficient vehicle tracking algorithm. A vehicle tracking framework based on multi-module fusion is proposed. The main work of this paper is as follows: firstly, a method of vehicle detection based on depth neural network is proposed. Based on the information of vehicle and its surrounding environment, and based on the network structure composed of candidate network and detection network, the robust vehicle detector is obtained by using the deep learning framework (Caffe,) training. Experiments show that the algorithm has achieved good results in various weather and traffic scenarios. Secondly, a model updating method based on incremental learning is proposed to overcome the defects of kernel correlation filter tracking method in model updating. By establishing the snapshot set of the early tracking model and combining the model collected by the nearest neighbor frame, the tracking model containing the principal components of the early and current target information is constructed by incremental updating. The robustness and adaptability of incremental updating are proved by comparison with other model updating methods. Finally, on the basis of vehicle detection and correlation filter tracking algorithm, vehicle tracking algorithm is proposed. Vehicle detectors are used to initialize multi-scale correlation filtering tracking algorithms and to correct tracking failures. Correlation filter tracking algorithm is used to realize short time single vehicle tracking. By using the correlation method based on the confidence degree of trajectory segment, the detection and tracking modules are integrated, and a more complete vehicle trajectory is formed, and the vehicle tracking in complex traffic environment is realized. The experimental results show that the tracking framework achieves the balance between vehicle speed and accuracy, and achieves excellent results in the actual road surveillance video.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【相似文献】