摘要
传统的无人车避障系统多采用传感器融合技术,成本较高,操作复杂。基于搭载单目视觉传感器的无人车平台,设计了一种轻量级的卷积神经网络(CNN)模型(SPSENet)作为无人车的转向控制系统,为其避障提供技术支持,同时运用随机权重图像融合算法对AirSim仿真环境下采集的图像样本进行数据增强,用于提高障碍物被树荫遮挡情况下的检测效果。SPSENet模型对Mobile NetV3的结构进行了调整,在每个深度可分离卷积和池化模块后引入通道注意力机制,并由全卷积层来连接输出层,更好地实现了对无人车从图像到转向预测的端到端控制。实验结果表明,SPSENet与经典分类网络相比,参数显著减少,性能指标有所提升,具有较好的实用性。
The traditional obstacle avoidance system of unmanned vehicle mostly uses sensor fusion technology,which has high cost and complex operation.Based on the monocular vision sensor of unmanned vehicle,a lightweight convolutional neural network(CNN)model(SPSENet)is designed as the steering control system of unmanned vehicle to provide technical support for obstacle avoidance.The random weight image fusion algorithm is used to enhance the image collected in the AirSim simulation environment and improve the detection effect of obstacles under the shade.The SPSENet model adjusts the structure of MobileNetV3,adds the channel attention mechanism after the depthwise separable convolution and pooling module,and the output layer is connected by the full convolution layer,which better realizes the end-to-end control of unmanned vehicle from image to steering prediction.The experimental results show that,compared with the classical classification network,the SPSENet has significantly reduced parameters,and the performance indexes are improved,which also has better practicability significance.
作者
刘浩强
闫冬梅
张新宇
顾德英
沙晓鹏
LIU Hao-qiang;YAN Dong-mei;ZHANG Xin-yu;GU De-ying;SHA Xiao-peng(School of Control Engineering,Northeastern University at Qinhuangdao,Qinhuangdao 066004,China)
出处
《控制工程》
CSCD
北大核心
2023年第3期443-448,共6页
Control Engineering of China
基金
河北省高等学校科学研究重点项目(ZD2019305)
秦皇岛市科技计划资助项目(201901B013)。
关键词
无人车
转向控制
图像融合
深度可分离卷积
通道注意力机制
Unmanned vehicle
steering control
image fusion
depthwise separable convolution
channel attention mechanism
