Copyright © All rights reserved micansinfotech.com
SOFTWARE: ASP.NET | VB.NET | C#.NET | RAZOR MVC 4 ASP.NET | RESTful Web services
This paper presents a framework of accident prediction with a new perspective. First, the new framework of Chain of Road Traffic Incident (CRTI) is proposed, in which the observed vehicle movement features are viewed as road traffic system’s external “performance” that, in essence, reflect the internal “health states” (safety states) of the system at a specific time. A two-stage modeling procedure of CRTI is then proposed using scenario-based strategy: 1) a support vector machine is utilized to classify leaving lane scene versus remaining in lane scene and 2) Gaussian-mixture-based hidden Markov models are developed to recognize accident versus non-accident pattern CRTI given the classified scene. Moreover, the application procedure of the CRTI framework to online collision prediction is proposed. Finally, a simulation test of a typical vehicle collision scene based on PreScan platform is designed and carried out for model training and validation, and has shown promising results in accident prediction using the proposed framework. The CRTI framework could provide a new foundation for developing early warning/intervention strategies in driver assistance system under complex traffic environments.
A scenario-based modeling and application framework is proposed for CRTI development, as profound discrepancies may exist in CRTI evolution process among different accident scenarios that are unlikely to be captured by one HMM struc- ture. In the paper, a vehicle’s movementis first categorized into either a Remaining in Lane (RL) scene or a Leaving Lane (LL) scene based on the trajectory of the vehicle using Support Vector Machine (SVM), as a driver’s lateral and longitudinal driving behaviors are expected to be very different in essence
More sophisticated design of rules would be further explored and examined in our future work by conducting road tests for real world appli- cation. For each of the defined accident scenario, an accident pattern HMM and a non-accident pattern HMM could then be developed accounting for different movement characteristics caused by lateral and longitudinal driving behaviors as well as scenario-specific environment.
Available data (offline data collected from driving simulator, with more details) are divided into a training set (75% of whole data) and a testing set (25% of whole data). Data are first preprocessed (including feature variable selection and data extraction ), and then fed to the training of a two-stage CRTI model that includes a 1 st stage SVM-scene classification model and a 2 nd stage HMM-accident pattern recognition model (their detailed algorithms).
Specifically, a finite set of accident pattern HMMs (i = 1) and non-accident pattern HMMs (i = 0) are trained for each developed accident scenario respectively, which constitute a library of CRTI models that would serve the basis for online prediction. The performance of the developed two-stage CRTI model is finally assessed using testing data.
Once the CRTI model is properly trained offline, online real-time movement data of vehicles (which could be obtained in connected car context) could be fed to the model to decide the best matched scenario the vehicle is currently in using the SVM classifier (along with the simple rules discussed earlier) and the most likely CRTI pattern the vehicle has (an accident pattern vs. a non- accident pattern) based on the scenario-specific HMMs.
For real-world application, the recorded online movement data could also later be added to the offline database for improving the CRTI model.
T HE number of traffic accidents and resulting fatalities and injuries remain high in China and many other coun- tries . Roadway accidents are believed to be the integrated results of many factors, mainly in four aspects including driver, vehicle, roadway and environment. Physiological and psychological limitations of drivers, limited performance of vehicles, improper road alignment, as well as harsh weather- caused risks (such as poor visibility and slippery road) may altogether lead to a roadway accident . To improve roadway safety, earlier works mainly focused on “passive” safety measures (to reduce injury or property damage) by improving the performance of vehicles; whereas recent works conducted more research on “active” safety strategies (to real- ize collision avoidance, such as pre-crash and auto-brake systems) by actively identifying drivers’ behaviors as well as perceiving surrounding (vehicles, roadway and environment) conditions.
Algorithms proposed in most previous works for collision avoidance are based on relative velocity and relative dis- tance (such as following distance for longitudinal safety, and time to lane cross for lateral safety) by utilizing vehicle kine- matics and dynamics. Although such metrics are easy to obtain and understand in decision making development, relationship among driver behaviors, vehicle states, roadway conditions, environment characteristics, and collision are dif- ficult to be fully described based on such metrics. Hence, several more advanced algorithms considering more risk fac- tors have been studied based on mathematical and physical methods
the new framework of Chain of Road Traffic Incident (CRTI) is proposed, in which the observed vehicle movement features are viewed as road traffic system’s external “performance” that, in essence, reflect the internal “health states” (safety states) of the system at a specific time.
A two-stage modeling procedure of CRTI is then proposed using scenario-based strategy: 1) a support vector machine is utilized to classify leaving lane scene versus remaining in lane scene and 2) Gaussian-mixture-based hidden Markov models are developed to recognize accident versus non-accident pattern CRTI given the classified scene. Moreover, the application procedure of the CRTI framework to online collision prediction is proposed
The paper presents a new framework of accident prediction on the basis of the proposed concept Chain of Road Traffic Incident (CRTI). The concept makes use of assumptions in the theory of HMM, where observed vehicle movement features are viewed as road traffic system’s external “performance” at a specific time, which in essence reflect the internal “health states” (safety states) of the system. A scenario-based two-stage modeling procedure of CRTI is constructed, includ- ing 1) the 1 st stage scene classification: SVM is utilized to classify Leaving Lane (LL) scene versus Remaining in Lane (RL) scene; and 2) the 2 nd stage accident pattern recognition: Gaussian-mixture based HMMs are developed to recognize accident vs. non-accident pattern CRTI given the classified scene.
Moreover, the application procedure of the CRTI framework to online collision prediction is proposed. Finally simulation of a typical vehicle collision scene is designed and carried out based on PreScan platform, with the recorded driving data then applied to the proposed two-stage CRTI framework. Promising results have been obtained in accident prediction using the proposed framework based on simulation.