基于信息融合的智能车障碍物检测方法

Obstacle detection method based on fusion information

针对单一传感器在智能车环境感知中的局限性,提出一种利用激光雷达融合相机进行智能车前方障碍物检测的方法。首先通过激光雷达获取目标的位置和速度,利用深度学习的更快的区域卷积神经网络(Faster RCNN)算法模型,训练实车采集数据,并用以检测图像中行人和车辆目标,获取单帧下各个传感器的目标检测数据。然后通过激光雷达与相机之间的标定,实现目标的三维数据的图像投影,利用改进的迭代最近点匹配(ICP)算法对其和图像检测目标点作匹配,并结合多帧数据,实现了目标的图像检测数据和激光雷达检测数据的融合。实车实验结果表明:该算法对车辆行人的识别率达到85%以上,在车辆方面提升了2.6%,行人方面提升了13.6%;算法的实时性比较好,每帧近似为84.89 ms。该算法能够较好地融合两个传感器采集的数据,使得获取的障碍物信息更为全面,能够用于无人车进行车辆行人障碍物实时检测,为智能车的局部路径规划提供决策依据。

For the limitations of single sensor in intelligent vehicle environment perception, an obstacle detection fusion method was proposed, combined with information from lidar and camera for the unmanned vehicles. Lidar could obtain position and velocity of targets. The pedestrian and vehicle targets could be detected by Faster R（ Region）-CNN（ Convolutional Neural Network） algorithm model from the image. Data from lidar could be calibrated. The improved ICP（ Iterative Closest Point）algorithm was adopted for registration, to achieve better data fusion between the key points of targets in image detection and target of lidar points. With considering multi-frames data from sensors, the fusion of image detection data and laser detection data could be achieved. The experiments proved that the recognition rate of pedestrians and vehicles reached more than 85%,with the enhancement of 2. 6% in vehicles and 13. 6% in pedestrians. Also, the algorithm has better real-time performance.The consuming time of algorithm for each frame was approximately 84. 89 ms. The algorithm can better combine the data between the two sensors, and can obtain obstacle information which is more comprehensive. It can be used to real-time detection of vehicle and pedestrian obstacles, and can help make planning for the local path in intelligent vehicles.