Context
Irider is an innovative company that has developed IRAS™, the first Intelligent Rider Assistant System. The IRAS™ software platform leverages artificial intelligence, computer vision, connectivity, and robotics to provide various solutions. Irider's solutions are specifically designed for manufacturers of light electric vehicles and operators of shared micromobility services, thereby contributing to enhanced safety and efficiency in the lightweight mobility sector.
Some individuals drive dangerously due to a lack of training, particularly scooter and free-floating bike riders. Others engage in risky driving behavior because of their working conditions, such as gig workers (Uber Eats, etc.). These behaviors lead to incidents and increase the insurance costs for all two-wheeler users, including those who adhere to traffic regulations.
As part of the development of a smart bike, we were tasked with implementing real-time roadway recognition through an onboard camera.
The Challenge
We needed to develop an embedded model that was lightweight enough to be integrated into a light electric vehicle (LEV), while also achieving near-real-time response times (less than 25 ms). Additionally, the model had to reliably detect the lane being used by the rider and distinguish between bike lanes, sidewalks, roads, bus lanes, and cycle tracks.
The Solution
We developed a segmentation model for bike lanes, car lanes, bus lanes, and sidewalks. The model was optimized for deployment on a TPU system (Google Coral TPU).
Results
The model was put into production in just two months, saving over €2000 on image labeling. It achieved a throughput of more than 25 images per second on an embedded system for instance segmentation.
