Research On Lightweight Visual Perception Methods In Complex Traffic Scenes

Complex traffic scenes are rich in components and diversified in their composition.The environmental perception technology of smart vehicles can not only help them accurately perceive and understand the surrounding situation,grasp their own driving status in time and interact with the external environment information,but also contribute to safety performance improvement.With the vigorous development of multimedia technology based on images and videos,the scene information obtained by visual sensing equipment is more abundant,reliable and at a lower cost.Therefore,vision-based environmental perception has become an indispensable way for intelligent vehicles to realize path planning,decision-making control and other functions.The lightweight perception method is to further reduce the complexity of the model and the amount of parameters on the basis of maintaining the accuracy,aiming to reduce the burden of computer storage and computation.It covers the idea of preserving the core of model architecture and reducing redundancy,as well as exploring the corresponding computational efficiency.This type of method has promoted the application of machine vision,control science and other technologies in mobile and embedded devices,and has been widely used in plenty of fields such as assisted driving and intelligent transportation.This article focuses on how to use image processing methods to accomplish more accurate environmental perception,using and designing novel lightweight perception structural models to help autonomous driving technology to realize the perception of high-resolution road images on a more portable and easy-to-operate processing platform,and improving the judgment and auxiliary control capabilities of smart vehicles for complex traffic scenarios.The crucial challenge of environmental perception comes from the system's sensitive interpretation and judgment of high-level visual content,which assists intelligent vehicles to quickly perceive dynamic information such as spatial layout and target characteristics in traffic scenes,and provides them with reliable and accurate response operations and basis for the execution.However,the road foreground and background in the actual traffic scene are changing all the time.A variety of complex targets have diversed features and are related to each other.Even different entities of the same object may cause scale changes or occlusions due to shooting angles and distances.Numerous objective factors have brought severe tests to the vision-based environmental perception methods,and the realization of global accurate understanding and information interaction still needs to be consolidated and improved.Accordingly,this article focuses on in-depth research and innovation on the key technologies of visual perception tasks in complex traffic scenes,uses machine vision-based ideas to perceive different levels of visual features,and constructs a few lightweight models and methods to improve environmental perception performance,and improve their precision and efficiency.First,the light-weight machine learning segmentation method is used to obtain the shallow features of images,that is,to complete the recognition and extraction of the drivable area in the road environment perception.Then,in order to understand the semantic content of the road more comprehensively,combined with the lightweight cascaded convolutional neural network to obtain deeper image features,to realize the scene semantic segmentation as the core solution to the environmental perception problem.Afterwards,in order to further realize timely perception and avoidance of vehicle obstacles that may exist in the scene,a lightweight vehicle target detection model is constructed.The research results have practical significance and reference value for the environmental perception and development of intelligent vehicles.The specific research content and contributions of this paper are summarized as follows:(1)Due to the diversity of structure,the complexity of texture changes,and the instability of natural exposure in road scenes,most conventional road segmentation-based detection methods have problems such as information redundancy,boundary loss,and blurring.This paper proposes a hybrid image segmentation model based on MS-RG and applies it to road image segmentation to extract the target travelable area.It can provide more position pixel information in an unsupervised manner and does not depend on the amount of data.Meanwhile,the image transformation enhancement theory is applied for preprocessing,the superpixel algorithm is utilized for more direct and efficient segmentation,and the region growth segmentation method for multiple seed points is optimized.Consequently,the problem of incomplete segmentation is solved.While the global and local pixel features are effectively processed,noise and grayscale unevenness are better controlled,so that the boundary information between each region of the image is segmented more clearly.It also effectively improves the robustness and real-time performance of the segmentation method.The experimental results on the road image dataset manifest that the model proposed in this paper effectively enhances the segmentation accuracy and real-time performance compared with other methods,is capable of identifying road information precisely in images,has strong applicability to meet actual application requirements.(2)Since deep learning is accurate in cognition of street environment,convolutional neural network has made great progress in the application of traffic road scenes.Nevertheless,complex network depth,large datasets,and real-time requirements are typical problems that urgently need to be solved to realize intelligent driving technology.This paper proposes an improved lightweight real-time semantic segmentation network,adopting the image cascade network architecture,while considering road semantic modeling and in-depth understanding of the road environment at the semantic level,and using multi-scale branching and cascading feature fusion unit to extract rich multi-layer features.In this paper,the spatial information network is designed to transmit more prior knowledge about spatial location and edge information.During the training phase,this paper also adds an additional weighted loss function to optimize the robust problem of road image semantic segmentation to enhance the learning process of the deep learning network system.This kind of lightweight network can quickly perceive the road semantic information and realize the environmental perception of the road scene according to the segmentation result,thereby meeting the needs of assisted driving.The experimental results on the road dataset indicate that compared with other popular semantic segmentation algorithms,the model in this paper has a significant improvement in the processing ability and timeliness of detailed information.(3)Carrying out vehicle obstacle detection tasks in complex road scenes,vehicle targets are susceptible to interference from scale changes between different object instances and occlusion by other traffic targets.The feature attributes extracted by the model may be inconsistent,especially the target objects and small-scale objects are prone to missed detection and misdetection.For the purpose of balancing detection performance and real-time requirements,this paper proposes a vehicle detection model based on multi-view feature fusion.A single-stage multi-frame target detection framework is adopted,and different task modules are used to mine the deep-level correlation information of the image to reduce the loss of local features.And it aims to ensure the efficiency of target obstacle detection in complex traffic scenes.For deep features,a feature fusion module is built to facilitate the algorithm's transfer of contextual information to multi-layer feature layers;for shallow features,multi-view modules and multi-scale features based on multi-branch convolution and expanded convolution are used to build feature pyramids,learning the location and type of obstacles in the target vehicle.The experimental results demonstrate that,under the premise of ensuring real-time performance,the model in this paper not only enhances the detection performance of multi-scale and obstructed objects in complex traffic targets,but also improves the efficiency of the environment perception algorithm.