基于Concenter-Net神经网络的无人驾驶汽车实时规划方法  

Real Time Planning Method of Driverless Vehicle Based on Concenter-Net Neural Network

作　　者：蒋荣军 JIANG Rong-jun(Department of Information Technology,Guangzhou Sontan Vocational College,Guangzhou 511370,China)

机构地区：[1]广州松田职业学院信息工程学院,广东广州511370

出　　处：《数学的实践与认识》2023年第5期164-171,共8页Mathematics in Practice and Theory

基　　金：教育部哲学社会科学研究2009年度重大攻关项目《义务教育学校布局问题研究》(CX123456)。

摘　　要：无人驾驶汽车在机器视觉领域中成为研究热点,提出一种高效的多尺度目标检测网络,称为ConcenterNet无人驾驶方法,方法可以定位远距离区域在图像中实时并集中于前方行车记录仪透视图中的远处物体,它可以检测消失点并集中在近距离区域.第一次物体检测模型的推理将产生更大范围的感受野检测结果和更高的分辨率预测可能的消失点位置,即帧中最远的位置.然后,将其发送到对象检测模型进行第二次检测推断,获得远距离目标检测结果.最后,将这两个结果与特定的非最大抑制(NMS)方法合并.所提出的网络架构可以应用于大多数对象检测模型,仿真结果表明,所提出的控制策略可以确保极端情况下,Concenter-Net架构模型使用较低分辨率输入大小,模型复杂度较低,可以显著提高精度和召回率,行驶稳定性和轨迹跟踪精度得到提高.The driverless vehicle has become a research hotspot in the field of machine vision.This paper proposes an efficient multi-scale object detection network,called the Center Net driverless method.This method can locate remote objects in the remote area in the image in real time and focus on the perspective of the front dash cam.It can detect vanishing points and focus on the near area.The reasoning of the first object detection model will produce a wider range of receptive field detection results and a higher resolution to predict the possible vanishing point position,that is,the farthest position in the frame.Then,we send it to the object detection model for the second detection inference to obtain the remote target detection results.Finally,these two results are combined with a specific non maximum suppression(NMS)method.The proposed network architecture can be applied to most object detection models.The simulation results show that the proposed control strategy can ensure that in extreme cases,the Center Net architecture model uses a lower resolution input size and has a lower model complexity,which can significantly improve the accuracy and recall,driving stability and track tracking accuracy.

关 键 词：无人驾驶汽车 机器视觉 目标检测、深度学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] U463.6[机械工程—车辆工程]