Abstract:

Real-time detection of road vehicles is a hot topic in the field of computer vision. Aiming at the problems of low detection accuracy and slow speed of road vehicle detection algorithms, a road vehicle target detection method based on improved YOLOv3 is proposed. Firstly, a convolutional neural network with 30 convolutional layers is constructed by improving the Darknet53 skeleton network, which reduces the network cost and improves the detection speed. Secondly, according to the characteristics of road vehicle aspect ratio fixed, the [k]-means clustering method is used to select the anchor point prediction bounding box, which improves the detection speed and accuracy. The experimental results show that the proposed method has an average accuracy of 90.08% on the standard dataset KITTI, which is 0.47% higher than the traditional YOLOv3, and the detection speed reaches 76.04 f/s, which is obviously superior to the traditional YOLOv3 algorithm. At the same time, the proposed method is applied to the vehicle driving dynamic data set, which can realize real-time detection of road vehicles in the video.

Key words: vehicle detection, YOLOv3, Convolutional Neural Network （CNN）, Darknet53, [k]-means

摘要：

道路车辆实时检测是计算机视觉领域中的研究热点问题。针对道路车辆检测算法存在检测精度低、速度慢等问题，提出了一种基于改进YOLOv3的道路车辆目标检测方法。通过改进Darknet53骨架网络构建了有30个卷积层的卷积神经网络，在减少网络成本的同时提高了检测速度；根据道路车辆宽高比固定的特点，利用[k]-means聚类方法选取锚点预测边界框，提高了检测速度与精度。实验结果表明，提出的方法在标准数据集KITTI上的平均精度达到了90.08%，比传统的YOLOv3提高了0.47%，检测速度达到了76.04?f/s，明显优于传统的YOLOv3算法。同时将该方法应用于车辆行驶动态数据集，能够实现针对视频中道路车辆的实时检测。

关键词: 车辆检测, YOLOv3, 卷积神经网络, Darknet53, [k]-means