Computer vision for advanced rider assistance system
A computer vision-powered system for motodrivers with real-time alerts against road threats.
+10% Model accuracy
50% resource reduction
Migration to Hailo Edge AI
About the Client
A well-funded startup specializing in Advanced Rider Assistance Systems (ARAS) for motorcycle riders partnered with Yellow to enhance their existing solution using computer vision for real-time detection of potential accidents and road hazards. They had a proven solution, validated by major companies like Coca-Cola and IBM, but aimed to improve model accuracy and enhance resource efficiency.
Product Key Features
Explore the standout features of our product that make it unique and beneficial for users.
Enhanced driver safety
The solution was tested by thousands of motorcycle riders in both general and extreme weather conditions, so the ARAS safety is proven to be effective.
360° detection
AI-based computer vision helps the system analyze real-time footage and detect all possible threats around the vehicle.
Updated technologies
The ARAS uses computer vision, data analytics, and cloud technologies to provide the best user experience.
Non-intrusive alerts
With the help of timely and nondisturbing alerts, the rider has enough time to react and avoid threats.
Development Challenges and Solutions
How our team coped with a range of development challenges.
Enhancing safety monitoring
Challenge: We needed to improve the accuracy of the front collision prevention module’s safety monitoring.
Solution: We used the XGBoost model for the task. The new model is optimized for better threat detection: Compared to the previous version, it now operates with 40 carefully selected features, and the core accuracy metrics increased up to 10%.
Model optimization
Challenge: The client wanted to upgrade the Detection & Classification module for better efficiency.
Solution: We optimized detection and classification models within the pipeline with the help of transitioning to more advanced architectures and applying targeted optimizations. The latest versions now consume 50% fewer resources while maintaining performance metrics.
Migration to an edge AI solution
Challenge: We needed to help the client migrate their solution from Jetson platform to Hailo AI.
Solution: We assisted in model conversion and adaptation for computer vision tasks. Our team also participated in extensive testing to ensure high performance on the new hardware. The migration aimed at reducing power consumption while maintaining real-time processing capabilities. As a result, it boosted system efficiency enough to enable smooth deployment and high accuracy in resource-constrained environments.
How it Works
RiderDome is an AI-powered Advanced Rider Assistance System ARAS designed to enhance motorcycle safety for individual riders and fleets.
Each motorcycle is equipped with front and rear wide-angle cameras and an onboard NVIDIA Jetson device. This edge unit runs AI models locally to detect vehicles, lane changes, and collision risks in real time. When a threat is identified, the system alerts the rider immediately via visual signals, without relying on cloud connectivity.
Safety events and ride data are uploaded to a cloud-based platform. Fleet managers access a web portal to monitor alerts, review incident history, analyze rider behavior, and gain actionable insights across the entire fleet to improve safety and operational efficiency.
Tech Stack
Technologies and tools we used to improve the solution.
Python
C++
XGBoost
EvalML
Alert Classifier
ONNX
HailoRT CLI
Model Zoo
Results and Future Plans
We helped improve model accuracy and enhance resource efficiency for the ARAS system.
Experience the system in action during an actual traffic scenario
By enhancing model accuracy by 10%, optimizing resource usage by 50%, and successfully migrating the solution to a cutting- edge AI platform, we significantly boosted the Advanced Rider Assistance System’s performance and reliability. Our improvements have resulted in safer, faster, and more efficient real-time alerts, helping riders stay protected on the road under all conditions.
Planned future work includes refining detector and classifier models to boost accuracy, adapting the software to new hardware, and improving device response to avoid false alerts in borderline cases.