Studies On The Key Algorithms And Application System Of Traffic State For Expressway Network Via LDM~3

In the field of intelligent transportation system,the Advanced Traffic State System(ATSS)plays a vital role.It reflects the macroscopic traffic state of expressway network,the trends of traffic flow,the regional distribution of adverse weather and its impact on driving safety,and the overall service quality of expressway network.On the microscopic level,this system reflects information on each single incident,traffic flow parameters such as speed and road occupancy,vehicle classification,the severity of the weather and other real-time information.All discussed above play an important role in travel time estimation,driving path selection,timely processing of emergency events,congestion control,road network construction,and comprehensive analysis of historical data.Based on the provincial expressway network and foreign experience,this dissertation focuses on video-based vehicle classification,real-time traffic state estimation,short-term traffic flow prediction,traffic incident detection and speed limit control.Moreover,a real-time traffic state system via LDM3(Local Dynamic Map/Multimedia/Management)is designed and implemented through cross-platform information exchange.This system is compatible with the existing intelligent transportation systems.From massive traffic video data,traffic information is detected based on visual sensing for mining traffic incidents,adverse weather information,vehicle types and other information.A camera-invariant feature is extracted from vehicle images and the prevalent sparse classification is introduced for vehicle classification in Pan-Tilt-Zoom(PTZ)videos.Experimental results on highways with real traffic show that the proposed method could reduce the computational complexity.In traditional methods,the traffic state is individually judged based on a single variable,which is not comprehensive.To overcome the shortcomings of existing algorithms,a method of traffic state estimation based on clustering analysis is proposed.A "segment-road-network" model of network-level traffic congestion index is constructed,which allows for the quantification of traffic congestion on each layer.Based on the spatial-temporal correlation effect,this dissertation analyzes prevalent prediction methods for single-segment short-term traffic flow,including time series,Kalman,and neural network.To reveal the interactions of traffic flow in different sections,the Laplacian matrix is introduced to describe the network structure,following which a short term traffic flow prediction method using network constrained Lasso(Least absolute shrinkage and selection operator)is also proposed.The method predicts the traffic flow at different time points in future.Traffic flow features are influenced by traffic incidents.To reduce the influence and time-delay caused by traffic incidents,an automatic traffic incident detection algorithm based on support vector machine is studied.Furthermore,by taking into account the changes in visibility and road surface friction under adverse weather conditions,an integrated variable speed limit control method for dynamic black area is developed.This dissertation is derived from a number of provincial projects including"demonstration project of intelligent highways emergency command application based on cloud-computing","demonstration system for information detection,mining,aggregation,release and aided decision based on sensor network",and"demonstration project and cloud platform of provincial intelligent transportation information service".The system developed in this dissertation has important theoretical significance and practical value.The main contributions of this dissertation are summarized as follows:?A real-time traffic state system via LDM3 is designed and implemented for expressway network.In this system,the traffic state is estimated and predicted through information exchange from multiple platforms.? A short-term traffic flow prediction method using network-constrained Lasso is proposed and implemented.Experimental results verified that the proposed framework could achieve above 90%accuracy in the 30-minutes-ahead speed predictions.? A camera-invariant feature is extracted from vehicle images and the prevalent sparse classification is introduced for vehicle classification in PTZ videos,which achieves above 90%classification accuracy.? Several methods are developed and optimized.Those methods include a traffic state estimation based on clustering analysis,a "segment-road-network" model of network-level traffic congestion index calculation,an automatic traffic incident detection algorithm based on support vector machine,and an integrated variable speed limit control method for adverse weather conditions.