Research On Vehicle Scene Understanding Technology Based On Multi-task Deep Neural Network

In recent years,the development and transformation of the automobile industry is a true portrayal of technological progress,and intelligent travel modes are the future development direction.Vehicle scene understanding technology,as a prerequisite for measuring the level of vehicle intelligence and a "key step" to realize automatic driving,has attracted much attention in the field of assisted driving and even automatic driving.However,how to provide sufficient visual information for smart cars in the limited on-board computing space is still a challenging problem.Semantic segmentation and object detection are the two core tasks to achieve vehicle scene understanding.Thanks to the development of deep learning,algorithm models for single tasks have emerged one after another,but the visual information provided in the field of vehicle scene understanding is still relatively limited.Multi-task neural network has the functional characteristics of processing multiple task requirements at the same time,and can provide multi-dimensional environmental information that meets the needs of vehicle scene understanding without causing too much computational burden,providing a good solution for realizing vehicle scene understanding.The subject of this paper is based on the deep learning method to study the drivable area information and target detection information required by the vehicle scene understanding technology.Propose a vehicle scene understanding method based on multi-task deep neural network to provide algorithm model and theoretical support for vehicles to obtain multi-dimensional environmental information.The main research carried out in this paper is as follows:(1)Aiming at the trade-off between model segmentation accuracy and reasoning speed in semantic segmentation tasks,this paper constructs a real-time semantic segmentation network FCENet based on feature context encoding.Using separable convolution to guide the design of the context extraction module,the spatial feature information encoding of the image in the down-sampling process,effectively reducing the network parameters;By introducing an attention mechanism to connect different levels of coding information for feature post-processing,improve the segmentation accuracy of the network.Constructed a real-time semantic segmentation network FCENet with good accuracy and reasoning speed,and set up multiple sets of comparative experiments on the relevant segmentation data set to optimize the network structure,the experimental results verify that the segmentation network designed in this paper achieves an effective balance between segmentation accuracy and speed.(2)Aiming at the trade-off between detection accuracy and detection speed in object detection tasks,this paper constructs a traffic participant detection network TPDNet based on YOLO fusion ROI pooling.Through the principle analysis of two target detection algorithms based on candidate regions and end-to-end regression,it is established that the Encoder-Decoder architecture is used to build the algorithm framework.The context extraction module in FCENet is used as the coding part of the detection network,and combined with YOLO’s fast regression idea and ROI pooling as the decoding part of the network.The experimental results verify the effectiveness of the detection network designed in this paper on the accuracy and speed of object detection.(3)Synthesize the algorithm model of segmentation and detection tasks,study the multi-task joint algorithm,merge two different algorithms by sharing the feature extraction process,and propose a multi-task deep neural network that realizes the segmentation and detection tasks at the same time.The experimental results show that the multi-task deep neural network model proposed in this paper can efficiently provide segmentation and detection information,and obtain a faster inference speed on the basis of ensuring the accuracy of the model.