Based on the theory of deep belief network ( DBN) , a front vehicle detection method by using multi-source information is proposed in this paper. Firstly the joint calibration of millimeter-wave radar and video camera is conducted and the transformation relation between two sensor coordinate systems is determined. Then through the preprocessing of millimeter-wave radar data, the label classification of obstacles is accomplished, and the data of front vehicle objects and other types of obstacles are obtained. Next the data are trained by utilizing DBN and the preliminary identification of front vehicles is performed. Finally the verification windows for front vehicle i-dentification are obtained according to the statistical data on the width and height of common vehicles. Test results show that with the method proposed, the correct rate of front vehicle identification is 91. 2％ and the total processing time for single frame image is 37ms, effectively raising the real-time processing speed of the system, in particular, it has good detection results for the vehicles in adverse circumstances like overcast, dark night, light rain, fog or haze, meeting the requirements of accuracy and stability for assisted driving.%本文中以深度置信网络为理论基础,提出了一种多源信息的前方车辆检测方法.首先将毫米波雷达和摄像机进行联合标定,确定两个传感器坐标系之间的转化关系.然后通过对毫米波雷达数据进行预处理完成前方障碍物的标签分类,获得前方车辆目标和其他类障碍物的数据.接着利用深度置信网络对数据进行训练,完成前方车辆的初识别.最终根据常见车型宽度和高度的统计数据获得前方车辆识别的验证窗口.实验结果表明,采用所提出方法前方车辆识别的正确率为91.2％,单帧图像的总处理时间为37ms,有效地提高了系统实时处理速度,尤其对阴天、夜间、轻雨或雾霾等恶劣的道路环境中的车辆有良好的检测效果,能满足汽车辅助驾驶对于准确性和稳定性的要求.
展开▼
