交通标志识别系统是智能交通的一个重要组成部分，在驾驶辅助、道路维护、导航系统和自动驾驶等很多方面都有广泛的应用。交通标志的检测是整个识别系统的关键点，也是难点所在。本文研究如何设计快速、有效的算法，从图像和视频中检测交通标志，主要的研究工作和贡献包括：

提出一种级联的快速交通标志检测方法。该方法的一个特色是在预处理步骤中使用中层显著性有效和可靠地滤除背景窗口，并在级联分类中快速提取特征来拒绝更多的非标志窗口。在滑动窗口搜索过程中，该方法结合邻域尺度感知技术，在高分辨率图像（1360x800）上达到3~5帧/每秒的检测速度，是大部分当前最好方法的2~7 倍。与这些方法相比，该方法在限制性标志上达到相当的性能，在危险和指示性标志上适当牺牲了一些性能。

提出一种用于视频的快速交通标志检测方法。该方法在改进图像中快速检测算法HHVCas的基础上，采用跟踪算法进一步加速。对图像中的检测算法，通过一种无监督的方法优化级联分类结构的参数，获得了和当前最好算法相当的性能，同时保持了速度上的优势。通过对一些计算步骤进行并行化，并在视频中借助跟踪算法和局部搜索区域维持，检测算子获得了2.8倍的加速，达到了37帧/秒的检测速率。通过利用多帧检测结果的时空一致性，在高召回率的情况下，将检测算法的精度提升了4~6个百分点。

提出一种两级最小费用网络流数据关联的方法来融合多帧检测到的交通标志。通过对短时序列中检测到的和潜在的标志构建最小费用网络流，我们找到其中的轨迹片段，并进一步将这些轨迹片段连接成完整的轨迹，同时进行虚警滤除和潜在标志召回。为了处理轨迹的短暂中断，我们使用片段水平的关联连接轨迹片段，将它们视为基本单元来构建第二级费用流网络，最后对检测到的轨迹进行分析，召回部分丢失的检测来填补短暂的轨迹中断。实验证明了该方法的有效性。

 

Traffic sign recognition plays an important role in intelligent transport, and is widely used in many applications such as driving assistance, road maintenance, navigation system, automatic driving, etc. The detection of traffic sign is a key and challenging issue in traffic sign recognition system. This dissertation aims at how to design fast and accurate methods to detect traffic signs in images and videos. The main research and contributions include:

We propose a cascade method for fast traffic sign detection. The main feature of the method is that mid-level saliency test is used to efficiently and reliably eliminate background windows in the preprocessing step. It also adopts fast feature extraction in the subsequent stages for rejecting more non-sign windows. Combining with neighbor scale awareness in sliding window search, the proposed method runs at 3~5 fps for high resolution (1360x800) images, 2~7 times as fast as most state-of-the-art methods. Compared with them, the proposed method yields competitive performance on prohibitory signs while sacrificing performance moderately on danger and mandatory signs.

We propose a fast approach for traffic sign detection from video. It is based on the improved image-based detector HHVCas and further accelerated by tracking. For the image-based detector, by optimizing the parameters in the cascade using an unsupervised approach, it obtains competing performance compared with the state-of-the-art while keeping the speed advantage. By parallelizing some steps, using tracking algorithm and local search region maintenance as well, the detector achieves 2.8x speedup and runs 37 fps in video. It also obtains precision increase by 4~6% at high recalls when exploiting temporal and spatial coherence of results in multiple frames.

We present a two-stage association method based on min-cost network flow to fuse results in multiple frames for traffic sign detection. Sign trajectory fragments in short sequences are found by using the network flow graphs, which are constructed with both detected signs and potential signs. The trajectory fragments are further concatenated into long trajectories followed by false alarm reduction and potential sign recall. To handle brief trajectory interruption, we extend the concatenation of trajectory fragments by fragment-level association, involving in constructing a second stage network with trajectory fragments. The final candidate sign trajectories are analysed, where some small gaps of trajectories are filled by recalling missing detections. Experimental results show the effectiveness of the method.