基于多任务神经网络的智能网联汽车环境自动感知研究

Research on automatic perception of intelligent connected vehicles towards environment based on multi-task neural network

为提高智能网联汽车的驾驶体验感,基于卷积神经网络原理,在卷积神经网络底层插入双线性插值层,改进卷积神经网络池化层,构建智能网联汽车环境自动感知多任务神经网络。通过多任务神经网络编码器提取采集的汽车环境图像特征,作为语义分割与目标检测解码器的输入,执行多任务神经网络训练操作,采用训练好的神经网络输出智能网联汽车环境自动感知结果。在不同道路环境、光线强度及噪声强度下验证基于多任务网络的智能网联汽车环境自动感知性能。结果表明:该方法在不同道路环境下可实现智能网联汽车环境自动感知,不同光线条件下的目标检测效果较好,网络实用性与语义分割精度较高,能够满足实际智能网联汽车环境自动感知需求。

In order to improve the driving feeling of intelligent connected vehicles,the automatic sensing method of intelligent connected vehicles towards environment based on multi-task network is studied by inserting the bi-linear interpolation layer into the bottom layer of convolutional neural network and modifying the pool layer of convolutional neural network.The collected image features of vehicles′environment are extracted by the encoder of multi-task neural network,which is used as the input of semantic segmentation and target detection decoder.The trained neural network is used to output the automatic sensing results of intelligent connected vehicles towards environment.The experimental results show that this method can realize the automatic perception of the intelligent connected vehicles towards environment.The target detection effect is good under different lighting conditions,and the network practicability and semantic segmentation accuracy are high.It can meet the requirements of the automatic perception of the actual intelligent connected vehicle towards environment.