Vehicle Route Inference Method Based On Deep Reinforcement Learning With Surveillance Camera Data

Smart transportation is the most important part of smart city construction.To build smart transportation,China has deployed traffic vehicle monitoring systems in various cities.The current traffic monitoring system can not only monitor and track vehicles,but also effectively detect rich peripheral information such as vehicle speed,driving direction,and identification of license plate numbers.Data mining based on traffic monitoring information has important application value and practical significance in the fields of traffic congestion prediction,urban planning,traffic control,and air pollution assessment.However,such huge monitoring traffic data is rarely used for urban traffic and urban computing.Therefore,this paper made the following work for large-scale traffic monitoring data:First,we implement vehicle type mining and application analysis research on a provincial capital city surveillance traffic data.First,we define three types of vehicles that have important influence on urban traffic,i.e.,periodic personal vehicle,taxi,and public commuting bus.We also propose the corresponding mining method combining vehicle category definition with frequent sequence pattern mining algorithm for each type of vehicles.Experiments on 120 million vehicle records collected from 1,704 surveillance cameras in Jinan demonstrate the effectiveness of our proposed definitions and mining methods.Second,taking the residential community as an example,we use four cases to analyze the residents' traffic modes and their relationships with the distribution of surrounding Points of Interest(POIs).Moreover,we also explore the potential applications of the citywide traffic data incorporated with POIs in urban planning,demand forecasting and preferential recommending.Second,the deployment of traffic monitoring is far from enough compared to actual road intersections.Even at the location of monitoring deployment,due to certain failures,it may not be possible to accurately obtain vehicle information.Therefore,the vehicle trajectories collected by traffic monitoring are sparse and incomplete.However,most urban traffic management,analysis,and prediction technologies rely on complete trajectory information,so this paper proposes a vehicle-specific route inference algorithm based on deep reinforcement learning.First,it combines deep reinforcement learning with traffic simulation software SUMO to introduce expert knowledge in the field of transportation.Multiple factors such as the difference in driving time,the number of turns,the number of traffic lights,the number of road segments,etc.are considered to design the reward function.A vehicle route inference algorithm combining traffic simulation and reinforcement learning is implemented.After that,on the real Jinan traffic monitoring data set,the effectiveness of the proposed algorithm was evaluated through comprehensive experiments.