Research On Road Traffic Environment Perception Based On Vision

The research and application of computer vision and deep learning are rapidly developing,but the visual environment perception and its application in low-cost assisted driving still pose challenges.Road environment perception is the foundation and key of automatic assisted driving systems.It can help vehicles to comprehensively understand their traffic environment through sensors such as cameras or radars and software.The vision-based perception system is favored due to its lower application cost.Among them,traffic sign detection can provide road traffic information,reduce traffic violations,improve safety,and is an essential function for smart transportation and autonomous driving;road drivable area perception can provide road information for the fusion and decision-making layers of autonomous driving,and then achieve path planning and improve driving safety.Therefore,this thesis focuses on the two major functions of traffic sign detection and road drivable area segmentation in the visual perception system.The main work contents are as follows:Firstly,YOLOv7-tiny object detection algorithm is analyzed in depth,and improvements are made for traffic sign detection applications: 1)Extract features of small targets to improve the ability of network,the feature extraction network is strengthened at the second layer of the backbone network.Specifically,an up-sampling process of the feature map is added to the original enhanced feature extraction network to obtain a 160x160 feature map,which is fused with the third layer feature map of the backbone network to obtain a larger feature map for small object detection.At the same time,anchor box parameters for small targets are added,and a detection head for the second layer of the backbone network is added.2)Triple attention mechanism is added to the four effective features extracted by the backbone network to obtain rich cross-channel information.3)All Leaky Re LU activation functions in the network are replaced with Si LU activation functions,enabling the network to more effectively capture linear/nonlinear abstraction and improve network performance.Secondly,Deeplabv3 plus semantic segmentation algorithm is analyzed in depth,and the Xception backbone feature extraction network with high computational complexity is replaced with the lightweight Mobile Netv3 backbone feature extraction network to better apply to mobile devices.In addition,the GAM global attention mechanism is added to the Atrous Spatial Pyramid Pooling module to adaptively learn the relationship between spatial position and channel and increase the network’s attention to important features.The cross-entropy loss function in the model is modified to Focal loss function to reduce the impact of class imbalance in samples and further improve segmentation accuracy.For the above improved algorithm models,multiple different traffic scene datasets are used for training and testing.Experimental results prove their feasibility and advanced.