Research On Intelligent Road Collaborative Perception Method Based On Visual Fusion And Deep Learning

Urban road traffic system is a complex system,and the various elements in this system,included pedestrians,vehicles and roads influence each other.The perception ability of traffic participants is limited and only a part of the current state of road traffic could be known,therefore traffic participants’ action is misjudged,which caused traffic complex problems,such as traffic congestion and traffic pollution.The current Intelligent Traffic System(ITS),which is used to overcome traffic complex problems,existed the shortage of acquiring insufficient road traffic perception information.ITS can only obtain the macroscopic traffic information,and not able to acquire traffic elements’ microscopic state and operation information.That is why Collaboration Vehicle Infrastructure System(CVIS)was put through to solve the problem of obtaining micro information through traffic elements.The real-time information of pedestrians,vehicles and roads can interact with each other by using wireless network in the CVIS environment,so that every traffic participant is able to obtain the current comprehensive road information,and make the entire road traffic system operative with stable,efficient and safety,as well as solved those complex traffic problems.CVIS is required to sense and process the running state of vehicles,roads and pedestrians in real-time.It is far from being able to obtain and process such magnanimity information only by connected vehicles with autonomous perception abilities.Intelligent road can assist connected vehicles to solve the problem of information acquisition and processing.By using edge computing technologies,road site units(RSU)are able to sense,judge,inference and make decisions about pedestrians,vehicles on the road and its self-information immediately,and achieve the overall control of the road system.The current road perception methods mainly use non-visual modal sensors to obtain vehicles’ and pedestrians’ information.Compared to non-visual modal sensors have the shortage of obtaining only one or few types of physical information.Visual modal sensors can obtain abundant physical information at the same time.Both visual modal sensors and non-visual sensors have the limitation of limited sensing range.Cooperative sensing can improve sensor’s sensing range and provide a full range of road perception effectively.In this paper,the following research is conducted on the collaborative perception method of intelligent roads:(1)Analyzed the complexity of road traffic system,demonstrate the advantages of CVIS in solving complex traffic problems,studied the characteristics of traffic elements in typical CVIS scenarios,and clarified the role of CVIS in dealing the problems that difficult to solved by classical ITS;(2)Compared the advantages and limitations of visual modal sensors and non-visual modal sensors in obtaining traffic elements,and used visual model sensor to obtain traffic elements due to its superiority.For the main road element,vehicles,proceed with object detection,object recognition,object tracking and traffic flow count based on deep learning,and finished the perception of vehicle operating information.(3)In response to the limitation of the single visual sensor’s limited perception range,proposed the methods and theories of visual collaborative perception.Compared the differences between traditional stereo vision perception and panoramic vision perception,and studied matching vehicles in different vision field,the key to the implementation road panoramic visual theory.Raised a visual collaborative sensing method under non-geometric constraints,which used deep learning vehicle re-identification to achieve vehicle matching.This paper presents the research on the intelligent road collaborative perception method based on visual fusion and deep learning,carried out vehicle detection,tracking and counting experiments based on computer vision and deep learning,and proceed vehicle re identification experiment based on deep learning.It provides theoretical support for the traffic management departments to obtain vehicle information by combing different roadside sensors to achieve more comprehensive access to vehicle information and more effective traffic guidance.