ML/Computer Vision + Data Annotation for ADAS

We worked with one of the leading automotive manufacturers in Japan to provide an AI object detection and quality analysis solution, enabling Advanced Driver Assistance Systems (ADAS). Our client is a leading Japanese automotive company willing to enhance the ADAS system by introducing advanced AI object detection and quality analysis. Our main task was to provide an AI-based system able to correctly detect unknown objects in real-time and support improving vehicle safety by giving timely notifications to drivers. We had to create a high-quality dataset and train complex Machine Learning models to achieve this.

We developed comprehensive Data Annotation guidelines and built the necessary infrastructure to start data collection. With the CVAT tool, our annotators labeled over 1.5 million objects in more than 20 categories related to the road environment. The categories included vehicles, pedestrians, roads, road markings, and background. We used advanced labeling techniques, such as polylines and polygons, to get minute details of objects.

The annotated images were processed by Deep Learning Neural Network algorithms. Our Machine Learning engineers developed custom Python scripts and implemented MLOps practices throughout the training pipeline to further automate and optimize this work to improve accuracy and efficiency. This project also encompasses general object detection, camera association labeling, and qualitative analysis through continuous remote and local team collaboration.

• Extensive Data Annotation: Annotation of more than 1.5 million objects across 20+ classes with high precision at 98.5% and above accuracy.
• Advanced Labeling: Polylines, polygons, and segmentation were used in the previous relevant lane-specific attributes with great detail.
• Custom Python Scripts: The team developed and implemented custom Python scripts that cover the distribution of the output in an effective way and automated data processing workflows.
• Deep Neural Network Training: We integrated annotated images into DNN algorithms to strengthen ADAS’s object detection capability.
• MLOps Integration: Our team implemented MLOps tools and practices to ensure seamless model deployment and scalability.
• Real-time Image Processing Workflow: We designed a pipeline to support continuous data flow starting from image capture, going through manual labeling, and a real-time model update.
• Scalable Infrastructure: The solution was expanded to support additional tasks and models, building the basis for future AI-driven automotive innovations.

• Managed to successfully process over 21,000 images with over 1.5 million annotated objects.
• The accuracy rate is higher than 98.5%, enabling the client to improve object detection in ADAS.
• Enhanced vehicle capability to detect unfamiliar objects and issue a real-time alert to improve driver safety and build confidence.
• The models enable the client to integrate them into ADAS, improving driving safety and reliability.
• We delivered a scalable solution that allows the client to lay the foundation for further development and deployment of other AI object detection applications and driver assistance.