Research On Traffic Scene Understanding Method Based On Depth Information

In recent years,for the complex and ever-changing road driving environment,domestic and foreign research institutions have introduced a variety of environmental awareness algorithms,especially the application of convolutional neural networks to environmental awareness tasks,which is a breakthrough in this field.However,the complexity and variability of traffic scenes make the task of visual perception very difficult.The accuracy and real-time performance have not yet reached the real vehicle requirements.How to balance the two aspects is also the focus that has been widely concerned.It is still necessary to further explore effective ways to improve network performance.This paper takes convolutional neural network as the core and explores a network structure based on depth information to realize the vehicle’s perception of road traffic environment.Mainly around the following aspects:First of all,this paper proposes a network model based on the encoder-decoder structure,which greatly improves the real-time performance under the premise of ensuring network accuracy.On the basis of this,the color map and depth information are fused,and the effects of the first-end channel fusion and the layer-by-layer feature fusion on the semantic segmentation result are compared.The experimental results show that the accuracy obtained by the layer-by-layer feature fusion method is higher.Then,based on the analysis of context information extraction method,an improved spatial pyramid pooling module is proposed for the problem that the segmentation result based on the encoding-decoding structure is rough.The module can fully extract dense image features of different scales and improve network performance.At the same time,we continue to explore the fusion method of color map and depth information,and propose an asymmetric parallel bifurcated network structure to extract the color map and disparity map information separately.The information is fused at the end of the network and eventually gets output.Finally,due to the complex network model,which consumes a large amount of memory and easily leads to the slow running speed of the network,the selected network is speed-up processed,and the model is compressed to further improve its real-time performance,so that it can be applied to the intelligent vehicle environment sensing system.This work provides a new idea for intelligent vehicles to understand the scene in complex traffic environment...