Research On The Method Of Freeway Accident Detection Based On Trajectory

With the increase of vehicle population,road traffic accidents are becoming more frequent,causing a large number of casualties every year.Especially on highways,traffic fatalities are high and secondary accidents are more likely to occur.In recent decades,more enterprises and scholars have paid attention to traffic accident detection and classification,mainly based on downstream traffic flow changes,single-vehicle gesture data and video data for accident detection,but these methods are not real-time,low reliability or difficult to obtain data.Therefore,it is of great significance to explore a fast and reliable detection method to shorten the rescue time of the injured and reduce the property loss caused by the accident,and as well as to provide theoretical support for the judgement of traffic accident liability in the future.At present,there have been a lot of researches on accident detection,however they are still not deep enough from the perspective of accident classification.The existing studies on traffic accident detection can be categorized into the three types from the viewpoint of the used data,namely,the method based on the macroscopic traffic flow data,the method based on the single-vehicle gesture data,and the method based on the video data.However,these methods all have some limitations.The real-time performance of accident detection based on macroscopic traffic flow data is low,and the reliability of accident detection based on singlevehicle gesture data is not high.The traffic accident research based on image has certain requirements on outdoor conditions such as weather and light,and the detection cost is high.With the popularity of Global Navigation Satellite System(GNSS)on mobile phones and onboard equipment,increasing amounts of real-time vehicle trajectory data can be obtained more and more easily,which provides a potential way to use the multi-vehicle trajectory data to detect an accident on freeways.However,a large number of accident experiments are impossible to be conducted in a field,besides,the cost of obtaining crash tracks through real car experiments are relatively high,the feasibility is low,so the trajectory data is simulated based on PC-Crash in the paper.PC-Crash is a professional traffic accident simulation software,which can produce the vehicles’ trajectories before,in,and after an accident based on the embedded vehicle dynamics models.Deep convolutional neural network(DCNN)can easily recognize and classify multi-feature data.Therefore,an accident detection and classification method based on multi-vehicle trajectory data and deep learning are proposed in this paper.Because the GNSS equipment has positioning errors of 5-20 meters,the false positive rate of the method by directly using the gap/distance received from GNSS between the two adjacent vehicles to judge whether an accident happens is very high.Different from the existing methods,the idea of the proposed method in this paper is to detect and classify traffic accidents according to the trajectory information of the two accident vehicles in a period of time.The proposed method not only uses the position information of the related vehicles but also tries to capture the development tendencies of the trajectories of accident vehicles over a period of time.In this paper,a DCNN model based on disaggregate accident vehicle trajectory is firstly established to detect accidents and identify accident types from normal driving.In order to obtain the accident trajectory(that is,disaggregated data),the simulated trajectory data was generated from PC-Crash,including the normal driving trajectory and the trajectory before,during and after the accident.In addition,six types of traffic accident for training are considered,and the influence factors of trajectory data such as frequency and duration are analyzed to verify the effectiveness of the model.Secondly,considering that the accident will affect the trajectory of other vehicles in the lane,this paper establishes a DCNN model based on the aggregate trajectory of vehicles.In order to obtain the trajectory data of surrounding accident vehicles,the trajectory of accident vehicles generated from PC-Crash was imported into the micro traffic simulation software SUMO and the trajectory of surrounding vehicles was automatically generated by combining with Python for simulation.Aggregating the generated multiple vehicle trajectory,the aggregating data is obtained,two types of traffic accidents are considered,and the influence factors such as the sampling number,data frequency and duration are analyzed to verify the feasibility of the model.The results show that,the highway accident detection method based on trajectory and DCNN can effectively detect and classify traffic accidents.In order to ensure the robustness of the model,the frequency and duration of trajectory data are analyzed.it can be found that the higher the frequency and duration of the trajectory data were,the higher the detection accuracy is,and lane change accidents are easier to be detected than rear-end collisions.In addition,for the accident detection model based on the aggregate vehicle trajectory,The higher the number of trajectory samples,the higher the detection accuracy.The model has stable and high accident detection and classification performances for the datasets with different frequencies and durations.Even for some challenging accident types,the proposed model is also robust and adaptive.Therefore,the feasibility and applicability of this accident detection model is high,which is helpful to accurately and quickly detect accidents,identify accident types,and judge the accident liability in the future.