Overview

VTLAS is the home of a Stochastic Simulation Platform for Assessing Safety Performance of Autonomous Vehicles in Winter Seasons.  The work is sponsored by the NSF OAC Core Program (Award Page). The project aims to develop an advanced cyberinfrastructure toolkit that significantly improves simulation algorithms and enhances fundamental knowledge of computing. The focus of the project is to create a stochastic simulation platform capable of thoroughly evaluating the capabilities of Autonomous Vehicles' (AVs) Automated Driving Systems (ADS), especially under adverse winter driving conditions. The motivation behind this research is to build a reliable tool that can accurately model stochastic vehicle behaviors, study vehicle dynamics, and predict potential AV safety risks when faced with icy or snowy road conditions. This is crucial for ensuring the safety and reliability of AVs before widespread implementation.

Key Objectives

• Physics-Regularized Modeling: Leverage the principles of physics through a Physics Regularized Gaussian Process (PRGP) model to enhance the machine learning process used for simulating vehicle interactions on icy/snowy roads.

• Crash Risk Prediction: Predict both multi-vehicle and single-vehicle crash probabilities in mixed traffic environments by integrating the traffic simulation model with a new vehicle dynamics model, accounting for the complexities of winter driving conditions.

• Safety Assessment: Conduct comprehensive safety assessments of AV performance on icy and snowy pavement by analyzing stochastic vehicle motions and the corresponding risk factors.

• Open-Source Platform Development: Integrate the developed models into an open-source software package with extensive documentation and numerous application cases. The goal is to create a public, cloud-based platform that is easily accessible and capable of incorporating new data streams for continuous model improvement.

Key Modules

The simulation platform consists of both frontend and backend modules. The frontend is composed of three main components. Data Importer guides users through the process of uploading roadside video footage (if available) and roadway information. It verifies data format and consistency, processes and converts the input into training datasets, and stores them in the backend database. Configuration allows users to integrate their own ADS control algorithms for testing. It also enables the customization of simulation parameters such as road geometry, pavement condition, weather conditions, traffic volumes, and more. To enhance usability, the platform offers robust default settings and pre-defined scenarios, while allowing users full flexibility to modify all parameters. Result Display presents the outcomes of the AV safety assessment, including predicted crash probabilities, types of potential crash events, and identification of high-risk locations within the simulated roadway environment.

The backend module features an Inference Engine that integrates seamlessly with the database and leverages various computational resources available in the PI’s lab, including GPUs and server clusters, to execute simulation tasks efficiently. To initiate the Similator, users begin by uploading data via the Data Importer and configuring the simulation parameters through the Configuration Interface. The Inference Engine then retrieves the relevant data and settings from the backend Database and executes the simulation models. Upon completion, the results are stored in the database and delivered to the Result Display component for visualization. Users also have the option to download the full set of results for further analysis or downstream applications. 

Direct output of the algorithm

Frame ID, Object ID, Bounding Box Relative Coordinates (Pixels)

Issues and Solutions

• Multiple Detections: One vehicle was detected multiple times (e.g., both car and truck category). This could be eliminated by eliminating simultaneous detection.

• Re-Detections:  Vehicles of the same make and model could not be distinguished by the tracking algorithm. These vehicles are likely to be labeled as the “same” vehicle. We need to clear the “memory” of the tracking algorithm.

• Lost Tracking: If it was only for several frames, we can do some simple imputations.

• Congestion Handling: Right placement of the cameras.