Study On Intelligent Vehicle Scene Perception Method Based On Multi-tasking Network

The automotive industry is developing rapidly and is in an era of reform.Intelligentization has become the direction of future development.As the precondition and key technology to improve the level of automobiles' intelligence,intelligent driving environment perception technology has become the focus of current research.Based on the deep learning method,this paper proposes an efficient and accurate intelligent vehicle environment perception algorithm based on vehicle camera,which realizes the deeper understanding of traffic scene and lays an important theoretical and application foundation for the development of intelligent vehicles.Firstly,through the research and analysis of semantic segmentation network model,a semantic segmentation neural network structure with the concept of “cardinal” and multi-branch fusion is proposed.The neural network is divided into the encode and the decoder.Based on the PSPNet,we design the encoder structure to extract the features of the input images,and the extracted features are input to the decoder.The deconvolution and up-sampling operations are used to expand the size of the input feature maps to have the same size as the input images.The experimental results show that the proposed semantic segmentation network has good real-time and segmentation accuracy performance.Secondly,the previously proposed neural network applied to semantic segmentation which achieves better feature extraction effect is applied to the object detection task to achieve detection task.Aiming at the object detection method of Fast R-CNN,the encoder part of the proposed segmentation network is used to realize the feature extraction of the object detection task,and the extracted feature maps are generated into the region of interest.The regions are process by ROI pooling and then applied to fully connected layer to obtain the classification and regression output of the target.The experiment results show that the proposed network achieved better detection accuracy and real-time performance on the object detection task.Finally,this paper improves a multi-tasking network that simultaneously implements real-time semantic segmentation and object detection tasks under the traffic scene.The network consists of an encoder that extracts features from a shared convolutional and decoders that implements different tasks including semantic segmentation and object detection.The encoder network extracts feature of the input images based on the previously proposed segmentation and uses the feature maps as an input to the semantic segmentation and object detection decoders.By implementing separately training and jointly training of two networks,the experimental results show that the proposed multi-tasking network can achieve higher detection real-time performance without sacrificing detection accuracy in multi-class semantic segmentation and multi-object detection tasks of traffic scenes.