Machine Learning Powered Roadside Asset Extraction using LiDAR

Traditionally, road assets are monitored and inventory controlled getting direct access to each asset, which can be very time-consuming and requires numerous field recordings and trained personnel. With recent advancements in data collection technologies, vehicle-based data collection platforms can collect millions of data points from all spatial directions at highway speed per second. The big data incorporates LiDAR point clouds, 360o degree imagery, and Laser Crack Measurement Sensor data. This paper presents the development of an innovative advanced machine learning algorithm capable of extracting roadside assets including traffic signs, guardrails, line painting, and rumble strips from the big database. The machine learning process starts with training steps in which hundreds of thousands of training datasets are used and then tested against the testing dataset. Once the testing database has passed at 99% or more, the trained program is ready to detect that asset. The techniques used to train machine learning algorithms to extract signs from the LiDAR database are developed using unsupervised clustering algorithm followed by autoclassification using machine learning classifiers with imagery. A similar approach has been taken to identify other assets such as guardrails, rumble strips, and line paintings. This process has been able to successfully identify and classify traffic signs from highways as well as urban and rural roads. The developed machine learning algorithm is programmed in parallel and performed at typical highway operation speed. Additional information about the geometry and retro-reflectivity properties are other important features that are also calculated and reported by this algorithm. The developed algorithm has been in production phase in the British Columbia Ministry of Transportations Asset Management project, in more than 13000 km of highways and has been able to pass all quality assurance mechanisms. This paper outlines the steps followed to develop a roadside asset extraction machine learning algorithm from the LiDAR and imagery database, as well as present a sample of the resultant roadway traffic sign asset database.