摘要
设计了一种快速准确的算法,实现了环境复杂、样本缺少情况下实时车辆检测和车型识别,特别是对三轮车的识别。利用一种改进的卷积神经网络(convolutional neural network,CNN)快速学习车辆特征,采用微调、分段训练以及多层特征图结合的策略增强网络特征学习能力,在小样本下尽可能全面地学习目标特征。摒弃繁琐耗时的区域推荐算法和后分类算法,利用单个网络直接预测图片中目标车辆的位置和车型类别,大幅提高了算法性能。实验结果表明,采用Ge Force GTX 1080 GPU时,该算法对各类车型识别准确度较为平衡,平均检测准确率高达72.2%,每秒检测帧数46.57,在雨天、晴天、夜晚、强光和树荫等各种复杂场景下均有较好的适应性,适用于真实视频监控下智能交通系统精确实时的要求。
We design a quick and accurate algorithm to achieve the detection and recognition of vehicles, especially the tricycles, in the complex environments of lacking of samples. Firstly, an improved convolutional neural network is used to learn vehicle features rapidly, then many methods such as fine-tuning neural network, combining predictions from multiple feature maps and phased training are used to enhance network' s learning with a few samples. By eliminating the tedious and time-consuming regional recommendation algorithm and the post-classification algorithm, the position and category of the target vehicle in the image are directly predicted by using a single net- work, which greatly improves the performance of the algorithm. The experiment shows that when using GeForce GTX 1080 GPU,the ve- hicle recognition accuracy of the proposed algorithm is relatively balanced, with an average detection accuracy by 72.2% , and the number of frames per second is 46.57. It owns better adaptability in all kinds of complicated scenarios such as rainy day, sunny day, night, light and shade and so on, which is suitable for the precisely real-time requirements of intelligent transportation system under the real video monitoring.
作者
吴玉枝
吴志红
熊运余
WU Yu-zhi;WU Zhi-hong;XIONG Yun-yu(School of Computer Science,Sichuan University,Chengdu 610064,China;Institute of Image and Graphics,Sichuan University,Chengdu 610064,China)
出处
《计算机技术与发展》
2018年第6期1-6,共6页
Computer Technology and Development
基金
国家高技术研究发展计划(2015AA016405)
关键词
卷积神经网络
车辆检测
车型识别
多特征结合
分段训练
convolutional neural network
vehicle detection
vehicle recognition
multiple feature maps
phased training