Generating Synthetic Scenarios to Test an AI-Enabled Traffic Measurement System
Traffic is complex! It is the sum of road design, road user behavior, and their interactions. Understanding this dynamic is central to making the transport system efficient, safe, and sustainable – in line with the UN Sustainable Development Goals 11 and 3. Viscando offers AI-enabled data-driven solutions for traffic analysis, safety diagnostics, intelligent traffic control, naturalistic data collection, and extended perception at the very core of smart cities and autonomous driving.
Software quality assurance must evolve as systems increasingly rely on supervised learning. No longer is all logic expressed by programmers in source code instructions. Simulation-based testing is an active research topic used to identify critical input efficiently. Moreover, the testing results can also guide subsequent training data collection campaigns – or possibly even complement training sets with synthetic imagery.
RISE is Sweden's research institute and innovation partner. Through international collaboration programs with industry, academia, and the public sector, they support the competitiveness of the Swedish business community on an international level and contribute to a sustainable society. The institute has 2,800 employees who are engage in and support all types of innovation processes. RISE is an independent, state-owned research institute with unique expertise and over 100 testbeds and demonstration environments for future-proof technologies, products, and services.
Viscando works closely with cities and the automotive industry to accelerate the digital transformation of the transport network. They provide key enabling insights from AI-enabled traffic measurement systems. Detailed movement data for individual road users, e.g., pedestrians, bicyclists, and vehicles, are analyzed enabling a plethora of smart city and autonomous driving insights and applications. Some of these are analyses of mobility bottlenecks and traffic safety issues, critical scenarios, human behaviour models, and extended perception for autonomous and automated vehicles, as well as assessment of safety implications from changes in the policies and infrastructure. Viscando's solutions are used in an increasing number of European countries and have also attracted wider international attention.
Fueled by Internet-scale data and enabled by massive compute, ML using Deep Neural Networks (DNN) has revolutionized computer vision. Viscando uses DNNs for object recognition and classification to identify and track distinct objects in traffic, e.g., cars, vans, pedestrians, and bicyclists. Successful DNNs tend to be trained on massive amounts of training data, i.e., video frames containing manually annotated objects of interest.
From a Quality Assurance (QA) perspective, developing systems based on DNNs constitutes a paradigm shift compared to conventional systems. No longer do human engineers explicitly express all logic in source code. Instead, hundreds of millions of parameter weights are trained on huge datasets. Standard software QA approaches such as code reviews and code coverage testing are less effective. Simulation-based testing is increasingly used to identify critical input to reveal system faults. Search-based software testing, i.e., using evolutionary algorithms or reinforcement learning to generate test cases that stress the limits of the systems, is an active research area within RISE. Automotive simulators, increasingly relying on modern game engines such as Unreal Engine and Unity3D, offer photorealistic virtual environments that can be used to efficiently and effectively test systems in a safe manner.
In this project, you will work with open-source simulators to create a digital model of a real-world intersection equipped with Viscando's solution. Inspired by real traffic scenarios measured at the intersection, you will create virtual counterparts in the simulator. Based on the virtual scenarios, you will identify a set of test parameters for which you will implement a search-based approach to generate specific test inputs corresponding to explicit test traffic scenarios. The goal is to support robustness testing of the traffic measurement system, i.e., investigate which parameter values are the most effective at provoking system faults. Candidate test parameters include occlusion, precipitation, illumination, and vehicle speeds. All of which are feasible to explore in a simulator.
The final scope of this project is decided in consultation with the supervisor at the university and with the contact person at RISE. Both the students and the client should be aware that a relevant academic underpinning is required in this project
Preferred: computer vision, machine learning, evolutionary computing, simulation-based testing.
0732-75 04 29