基于卷积神经网络的限速标志实时识别及警示系统研究

发布时间：2018-06-25 02:53

本文选题：限速标志 + 实时识别　；参考：《吉林大学》2017年硕士论文

【摘要】：限速标志是常见且重要的禁令标志,限速数字为驾驶人提供道路限制速度,引导驾驶人安全驾驶。但是驾驶人常会忽视道路上的限速标志,例如在开车时打手机、与乘客聊天或者限速标志处于视觉盲区,致使车辆行驶速度过快,造成交通违法,发生交通事故。因此,限速标志识别及警示系统的开发可有效地减少交通事故的发生,保障驾驶人的生命财产安全。本文基于卷积神经网络进行了车载限速标志实时识别及警示系统的开发,主要开展了以下几项工作:对国内外交通标志识别、深度学习及相关系统的研究现状进行了分析,对限速标志自身的形状、纹理特征及设置方式、设置位置特征进行了分析;然后对该系统潜在用户的需求、系统开发的技术难点进行了分析,对系统的操作方法及运行界面进行了介绍;通过分析现状,选择了相应的硬件设备配置进行系统开发;而后通过分析各自的特点和优势,选择Open CV作为系统开发的数据库,选择Hierarchical Data Format作为编程过程中进行图像处理的数据格式。进行图像预处理及限速标志定位。首先根据车辆的运行状态和驾驶人的视认特性对程序的处理频率进行了分析,并根据限速标志分布位置特征进行交通标志粗定位;然后基于HSV空间颜色模型进行颜色分割,而后进行形态学滤波,对图像进行预处理;接着结合长宽检测与霍夫圆变换检测方法进行了圆形标志精确定位,并利用限速标志的灰度分布及数字分布特征进行了伪目标排除,从而精确定位限速标志位置。基于卷积神经网络对限速标志图像进行了识别。首先对截取后的图像进行预处理,即对待检测图像进行图像增强和归一化处理两个阶段的处理,并通过对大量图像进行分析判断得出无需对限速标志进行倾斜校正;然后构建卷积神经网络模型,即Le Net-5模型,包括卷积核为5×5的卷积层、采用最大池化法的池化层、Dropout层及全连接层,激活函数为Re LU函数,运用Softmax分类器进行分类;接着选用GTSRB数据集及实车采集的部分图像作为数据集,基于Caffe深度学习框架进行训练和测试。对限速标志实时识别及警示系统进行了系统开发及道路试验性能测试。首先,对系统的开发过程进行了介绍;其次,基于Caffe深度学习框架对系统进行了训练与测试,并对其结果进行了分析总结,即分别分析训练学习率、测试准确率、损失函数与迭代次数的关系,在训练过程中,系统呈现较好的收敛性,并且在测试过程中达到97%的测试准确率;最后对系统进行性能测试,经过测试,该系统对于不同安卓设备具有很强的兼容性,并通过实车试验验证该系统的性能。在不同车速状态下,限速标志实时识别与警示系统的识别准确率均能够达到85%以上,最高可达94%,并且通过试验过程发现了其具有较强的实时性、系统稳定性以及识别鲁棒性。
[Abstract]:Speed limit sign is a common and important ban sign. The speed limit number provides the driver with road speed limit to guide the driver to drive safely. However, drivers often ignore the speed limit signs on the road, such as talking to passengers on their mobile phones while driving, chatting with passengers or having speed limit signs in the visual blind area, which causes vehicles to travel too fast, causing traffic violations and traffic accidents. Therefore, the development of speed limit sign identification and warning system can effectively reduce the occurrence of traffic accidents and ensure the safety of driver's life and property. In this paper, based on convolution neural network, the real time recognition and warning system of vehicle speed limit sign is developed. The main work is as follows: the research status of traffic sign recognition, depth learning and related systems at home and abroad are analyzed. This paper analyzes the shape, texture feature and setting mode of the speed limit sign, then analyzes the demand of the potential users of the system and the technical difficulties of the system development. This paper introduces the operation method and running interface of the system, selects the corresponding hardware equipment configuration to develop the system through analyzing the present situation, and then chooses Open CV as the database of the system development by analyzing their characteristics and advantages. Hierarchical data format is selected as the data format for image processing. Image preprocessing and speed limit location are carried out. Firstly, the processing frequency of the program is analyzed according to the running state of the vehicle and the recognition characteristic of the driver, and the rough location of the traffic sign is carried out according to the distribution feature of the speed limit sign, and then the color segmentation is carried out based on the HSV space color model. Then the morphological filter is used to preprocess the image, and then the circular mark is located accurately by combining the length and width detection with the Hough circle transform detection method, and the pseudo-target is excluded by using the grayscale and digital distribution characteristics of the speed limit sign. Thus accurate positioning speed limit mark position. Based on convolution neural network, the speed limit image is recognized. Firstly, the image is preprocessed, that is, image enhancement and normalization are processed to detect image, and through analyzing and judging a large number of images, it is concluded that there is no need for skew correction of the speed limit sign. Then a convolution neural network model, Le Net-5 model, is constructed, which includes convolution layer with convolution kernel of 5 脳 5, Dropout layer with maximum pool method and full connection layer, activation function with re LU function, and classification with Softmax classifier. Then GTSRB data set and some images collected by real vehicle are selected as data sets, which are trained and tested based on Caffe depth learning framework. The real-time recognition and warning system of speed limit sign is developed and the performance of road test is tested. First of all, the development process of the system is introduced. Secondly, the system is trained and tested based on Caffe depth learning framework, and the results are analyzed and summarized, that is, the training learning rate and the test accuracy are analyzed respectively. The relationship between the loss function and the number of iterations, in the process of training, the system shows good convergence, and in the process of testing, the accuracy of the test is 97%. Finally, the performance of the system is tested, and the system is tested. The system has strong compatibility with different Android devices, and the performance of the system is verified by a real car test. Under different speed conditions, the recognition accuracy of speed limit sign and warning system can reach more than 85%, and the highest recognition accuracy can reach 94%. Through the experiment, it is found that the system has strong real-time, system stability and recognition robustness.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;U463.6;U495

【相似文献】