Urban Traffic Flow Estimation Based On Visual Spatio-Temporal Information

Research on the visual based estimation methods of traffic flow parameters has always been one of the research focuses in intelligent transportation system(ITS).It is one of the key technologies to solve urban traffic problems.It has important theoretical significance and application value.The traffic flow estimation still faces many challenges in the application of complex urban traffic.1)In dense urban traffic scenes,vehicles are dense and there are occluded targets,small-size targets,which increase the difficulty of processing a single target in the traffic flow in traffic flow estimation;2)The urban traffic scene is complex,and the spatio-temporal correlation process based on a single target is complex,time-consuming and low in accuracy,which makes it difficult to estimate traffic flow effectively and in real time;3)The method based on traffic video can only output traffic flow parameter data,while the traffic flow estimation method based on the visual temporal spatial image(TSI)of traffic flow has some difficulties,such as dense target,different size and apparent deformation.These difficulties make it difficult to detect or segment the target in TSI.The above three problems have seriously affected the accuracy of visual based traffic flow estimation.This paper researches the vision-based traffic flow estimation algorithm for the above three challenges.The main content and contributions are summarized as follows:1.There are dense vehicles,occluded vehicles and small size vehicles in urban dense traffic.And these difficulties make the detection and tracking achieve low accuracy.Aiming at these existing difficulties,the traffic flow estimation based on pyramid multi-scale detection and restricted multi-targets tracking method is designed.The main contents and innovations are as follows:(1)The pyramid YOLO vehicle detection algorithm is proposed,and a pyramid multi-scale feature map is constructed through the scale variation mechanism.The multi-scale features are extracted by YOLO network and vehicle targets are detected.This improves the detection accuracy of occlusion vehicles,and small size vehicles.(2)W e propose the restricted multi-targets method based on line-crossing probability function.Based on the vehicle detection results,every vehicle that may cross the counting line is given a line-crossing probability value in each frame of the traffic video.And then some vehicles with higher probability values are selected to be tracked,so that the complexity of the tracking process is reduced,and the accuracy of the vehicle counting is improved.(3)Based on the pyramid YOLO detection and restricted multi-target tracking resluts,the estimation models of three traffic flow parameters are proposed.The traffic density is estimated according to the vehicle detection results.The traffic volume and speed are estimated according to the vehicle counting and tracking trajectory respectively.2.Traffic flow estimation still relies heavily on the spatio-temporal information correlation results of complex tracking methods,and it is difficult to obtain better accuracy and speed in complex traffic scenes.This paper proposes a bi-directional traffic flow estimation method based on STCF and cLSTM network.The main contents and innovations are as follows:(1)Designed for the whole traffic flow rather than a single vehicle,the STCF is proposed to represent the state of bi-directional traffic flow,which realizes the effective extraction of the spatio-temporal counting feature of the whole traffic flow.(2)A counting cLSTM network is proposed,which can effectively process STCF spatio-temporal characteristics and avoid the use of complex multi-target tracking methods.The process of spatio-temporal information processing is fast,accurate,and can adapt more scenes.(3)An accumulator with adaptive parameters is constructed to further process the output of the cLSTM network.The accumulator is designed based on an adaptive threshold to perform a secondary recognition of the traffic flow state,and then perform bi-directional vehicle counting and bi-directional traffic flow estimation.3.Traffic flow estimation methods bansed on traffic video only output traffic flow parameters,but cannot visualize the spatio-temporal state of traffic flow vehicles and estimate visualize traffic flow.To address the problem of visualized traffic flow estimation,we construct a visualized traffic flow:traffic flow temporal spatial image(TSI).The TSI contains spatio-temporal information of traffic flow.This paper proposes the spatio-temporal multi-scale deep regression network TM-Net,to extract spatio-temporal features from TSI and estimate the TSI density map,obtaining vehicle counts and visualized traffic flow estimation results.The main contents and innovations are as follows:(1)The automatic density map generation algorithm based on the TSI generation process is proposed,which solves the problem of obtaining data samples of the TSI density map.(2)Aiming to address the problem of different sizes,apparent deformation of the target and dense target in the TSI,we propose stacked multi-scale modules and the attention module in TM-Net.This can effectively estimates the TSI density map and achieves a relative high accuracy.(3)Based on the estimation results of the TSI density map,a visualized traffic flow estimation model is designed to realize the traffic flow estimation in an arbitrary time.