基于卷积神经网络和RGB-D图像的车辆检测算法

Vehicle Detection Algorithm Based on Convolutional Neural Network and RGB-D Images

针对利用彩色图像进行车辆检测时会受到路面阴影、车辆反光和光线不足等复杂情况影响的问题,提出一种基于卷积神经网络并融合彩色与深度图像的车辆检测算法。设计单通道RG-D融合网络和双通道RGB-D融合网络两种改进模型,分别用于提高检测速度和准确度。实验使用GTA(Grand Theft Auto)车辆数据集对该算法进行测试,并与基于RGB图像的其他流行算法进行对比和分析,结果表明:与基于彩色图像的Yolo v2算法相比,利用双通道RGB-D融合网络检测的准确率和召回率分别提升5.69%和6.31%,利用单通道RG-D融合网络对单一图像的最快检测速度达到24ms。实验证明,基于RGB-D图像的改进网络模型能够实现实时检测,并有效提高车辆检测精度。

Aiming at the problem that using RGB images for vehicle detection are affected by complex conditions such as road shadow, vehicle reflection and insufficient light. The paper proposes a vehicle detection algorithm based on convolutional neural network and combination of RGB and depth images. Two improved models of single-channel RG-D and double-channel RGB-D fusion networks are designed to improve detection speed and accuracy respectively. The algorithm is tested with (Grand Theft Auto) vehicle dataset and compared with other popular algorithms based on RGB images. The results show that compared with Yolo v2 algorithm based on RGB images, detection accuracy and recall rates increase 5. 69% and 6. 31% respectively by double-channel RGB-D fusion network, and the fastest detection speed of single image reaches 24ms with single-channel RG-D fusion network. Experiments show that the improved network model based on RGB-D images can achieve real-time detection and effectively improve vehicle detection accuracy.