Research On The Semantic Synthetic Technical Research Based On The Driving Scene

With the rapid development of the automotive industry and deep learning,autonomous vehicles have become a new research hotspot.An autonomous vehicle consists of an environment sensing system,a decision-making system and a control system.Among them,the environmental perception system is mainly responsible for collecting and processing the environmental information around the vehicle.Good environmental perception ability is an important guarantee for the driving safety of autonomous vehicles.Visual environment perception technology plays an important role in environment perception system,so it is of great significance to study visual environment perception algorithm.Based on deep learning method,this paper studies semantic segmentation technology in unmanned vision tasks,and proposes a real-time semantic segmentation model based on RGB color images,NASPNet.In the real-time semantic segmentation model NASPNet,first of all,this paper designs a low-level feature attention module to integrate the low-level features of the model and achieve better segmentation results with lower computational cost.Secondly,in order to obtain a larger receptive field without reducing the resolution of the feature map,NASP module is proposed.In addition,this paper also designed a multi-scale classification module to obtain different scale classification information to improve the segmentation effect.Experiments show that NASPNet can achieve the balance between segmentation accuracy and inference speed.Under good lighting conditions,RGB color images are usually obtained using an ordinary camera as input information to the model.However,in the case of poor lighting conditions,the semantic segmentation model proposed based on RGB color images often suffers from model degradation due to information loss in RGB image information obtained by ordinary cameras.It is an important means to solve model degradation to use data in different modes for information fusion.Compared with color cameras,infrared cameras obtain picture information through the thermal radiation characteristics of objects,does not depend on external light,and has all-weather characteristics.Therefore,this paper also studies the semantic segmentation task of RGB-T,and proposes a semantic segmentation model DSAMNet model based on RGB-T.In the process of semantic segmentation of RGB-T,this paper firstly proves the effectiveness of infrared image in improving model segmentation accuracy through experiments.In addition,focusing on the effective fusion of data of different modes,this paper proposes a differential space attention module based on spatial attention mechanism in multi-level feature fusion,so that features of different modes can be better fused,so as to improve the segmentation accuracy of the model.The NASPNet and DSAMNet models proposed in this paper achieve better results on Cityscapes and MF datasets,with mIoU of 76.2% and 54.4%,respectively.In order to obtain faster model reasoning speed,TensorRT is used to deploy NASPNet and DSAMNet.The test results show that NASPNet and DSAMNet can meet the needs of automatic driving in real-time perception.