基于注意力机制的实时车辆点云检测算法

Vehicle point cloud detection algorithm based on attention mechanism

针对现有激光点云目标检测效果、实时性差的问题,提出了一种基于注意力机制的实时车辆点云检测算法。本文所提出的检测算法将注意力机制算法与YOLOv3相结合,利用注意力机制对点云鸟瞰图的特征进行权重分配,以学习不同通道和空间下特征的相关性,并通过CIOU loss和Focal loss来改进检测器的损失函数。实验结果表明基于注意力机制的车辆点云检测算法检测速度可达30帧/秒,车辆目标的平均检测精度达到了92.5%。并且在实车数据测试中,该算法能快速准确的对一定范围内车辆进行准确识别,并且达到实时检测效果。

Aiming at the problem of poor detection effect and real-time performance of existing laser point cloud target,a real-time vehicle point cloud detection algorithm based on attention mechanism is proposed.The detection algorithm proposed in this paper combines the attention mechanism algorithm with YOLOv3,uses the attention mechanism to distribute the weight of the features of the bird′s-eye view of the point cloud,so as to learn the correlation of the features in different channels and spaces,and improves the loss function of the detector through CIOU loss and focal loss.The experimental results show that the detection speed of the vehicle point cloud detection algorithm based on the attention mechanism can reach 30 f/s,and the average detection accuracy of the vehicle target can reach 92.5%.In the real vehicle data test,the algorithm can quickly and accurately identify vehicles in a certain range,and achieve real-time detection effect.