毕业生知识库

    ITS（智能交通系统）是解决交通拥堵、交通安全和交通污染等城市交通问题的一种有效方法，能够提高交通运行效率、降低交通事故发生频率。交通视频监控系统作为 ITS 的重要组成部分，能够实时监控动态交通信息，为交通管控、交通诱导、交通规划等提供数据支持，受到研究者的广泛关注。随着视频监控设备的发展，高清摄像机在城市交通环境中得到广泛应用。高清图像具有更宽阔的视野、更复杂的背景以及更多的目标图像细节，这些特点为视频监控的研究提供了新的机遇与挑战。如何在百万级分辨率的高清图像中检测车辆信息成为众多学者关注的课题。基于视频图像的车辆检测和分类方法的研究是开展相关研究的基础性和关键性问题，具有巨大的发展潜力和广阔的应用前景。 

    针对交通监控系统实际应用的需求，本文从车辆检测算法的具体实现出发，研究了基于滑动窗扫描实现车辆检测中的实时性、检测率等问题。车辆检测和车辆分类技术是交通视频监控系统的基本任务。利用目标检测和分类任务间的关联性和互补性，本文将车辆检测和车辆分类任务进行了联合统一，提出能同时提供目标位置信息和类别信息的车辆信息识别框架。与此同时，本文也开展了深度学习关于车辆图像的特征表示和分类模型的研究，最终实现车型识别系统。本文的主要工作有： 

    (1) 针对城市交通场景中的环境噪声（车道护栏、花坛等），提出车辆检测模型和对称性信息互补的方法，实现从复杂道路环境中检测出多姿态、多类型和多尺度的车辆。该方法可以克服道路隔离护栏带来的误检问题，同时可以抑制部分遮挡带来的干扰。 

    (2) 针对滑动窗扫描的目标检测方法耗时长、效率低的缺点，提出基于混合特征的似物性采样方法。该方法利用贝叶斯框架融合了多尺度显著性、颜色对比度、边缘对比度等3种图像特征，能够提升检测效率、快速定位车辆，显著减少候选目标窗口的数目，快速、精确地确定车辆目标的位置先验。 

    (3) 提出一种车辆检测和分类技术融合的框架，为快速实现目标检测和分类任务提供了研究思路。不同于传统的利用二分类问题解决车辆检测的方法，该框架将车辆检测过程转化为预定位和多目标分类融合的过程，提出基于多分类的车辆检测和分类思路。该框架不需要人为的先验知识，可以同时获取车辆目标的位置信息和类别信息。 

    (4) 利用深度学习技术对车辆图像的特征表示和分类模型进行了探索，训练了车型识别的深度卷积网络模型。通过与似物性采样方法相结合，对提出的车辆检测和分类融合的框架进行了验证和具体实现，达到利用单一模型实现不同位置的小汽车、公交车、以及出租车的检测和分类的目标。 

    ITS (Intelligent Transportation Systems) is recognized as one of the most effective ways to solve traffic problems such as traffic congestion, traffic safety and traffic pollution. It aims to improve road efficiency and reduce traffic accidents. As an important component of ITS, Video Surveillance System can obtain dynamic traffic information in real time, and provide data for traffic management, traffic guidance, traffic planning and so on. Video Surveillance System has attracted much attention of researchers. With the development of video surveillance equipment, HD (High-Definition) cameras are widely used in urban traffic environments. HD images include a broader vision, complex backgrounds and objects details, which provides new opportunities and challenges for video surveillance. How to extract data from these images has become a hot topic. Vision-based vehicle detection and classification are fundamental and critical issues in Video Surveillance Systems and have huge develop potential and wide application domains.
    To meet the application requirements of traffic monitoring systems, this paper studies problems of processing efficiency and detection rate, and implements vehicle detection methods. The paper analyzes vehicle detection and vehicle classification techniques, and combines the two tasks as a coupled problem. By using the correlation and complementarity between them, a framework is proposed which can provide information about an object’s location and its category simultaneously. At the same time, this paper also do some research on how to use deep learning to represent a vehicle object and classify it. Finally, a vehicle recognition system based on deep learning techniques is realized. The main work of this paper includes the following aspects:
    (1)To deal with noise in urban traffic environment (Lane guardrail, flower beds etc.), this paper proposes a novel vehicle detection method based on Active basis model and symmetry information, which can locate different vehicles from complex environment that full of road boundary or fences. At the same time, this method can deal with part occlusion problem in some degree.
    (2)Existing sliding-window-based object detection methods are high-cost and low-efficiency. To improve it, a method of objectness proposals with hybrid features is proposed. This method can sharply reduce the number of candidate windows, and find prior positions of vehicles in images quickly and accurately. To improve detection efficiency, the method uses a Bayesian framework to integrate three image cues which include multiscale saliency, color contrast and edge density.
    (3)A framework of coupled vehicle detection and classification is proposed, which provides a new idea on how to combine basic tasks of object detection and classification. Traditionally, vehicle detection is viewed as a two classification problem. This paper transforms this task into a combinaion of vehicle proposals detection and multi classification. An idea of multi class-specified detection which combines vehicle detection and classification is proposed. The proposed framework requires no prior knowledge, and provides the location information and category information of a vehicle at the same time.
    (4)The paper also do some research on how to use deep learning to represent a vehicle object and classify it. A CNN (Convolutional Neural Network) model to represent vehicles has been trained, which obtains vehicle locatin and types with a single model. Combined with the objectness proposals method, the trained neural network can classify cars, buses, and taxis in different locations with a single model.