AI-Based ADAS (Advanced Driver Assistance Systems)
A computer vision model automatically identifies the type of road scene (city, highway, tunnel, etc.) and weather conditions (cloudy, rain, snow, etc.) from dashcam frames, reducing the need for manual annotation.
Tasks
- Develop a model to identify the scene and weather in dashcam frames.
- Achieve ≥ 90% accuracy on validation.
- Minimize false positives through post-processing (sliding window).
- Reduce the volume of manual annotation and accelerate dataset creation.
About the Project
Before the project, data quality was ensured by assessors who manually reviewed and classified millions of frames. This required significant resources and increased dataset preparation time.
An automated classification system was developed based on PyTorch and the ResNeXt/ResNet family of networks to quickly filter out incorrect frames and provide annotations in real-time with minimal human involvement.
Results
Challenges and Solutions
- Data Heterogeneity
- Frames varied significantly in lighting, time of day, and quality. Augmentations (color shifts, noise, cropping) and class balancing were applied to ensure model robustness.
- Ambiguous Class Boundaries
- “City” and “highway” scenes sometimes overlapped (e.g., suburban areas). Mixed categories were introduced, and annotation instructions were refined to improve class distinction.
- Sporadic Model Errors
- Individual frames could be misclassified due to glare or artifacts. A sliding window approach was implemented across frame sequences, with decisions based on the mode of predictions, eliminating most error spikes.
Related Services
Artificial Intelligence
We create innovative solutions based on artificial intelligence for your business.
Read More
Outsourcing
Our team of skilled developers is ready to create websites and applications of any complexity. We build modern web solutions using cutting-edge technologies and architectural approaches.
Read MoreReady to discuss your project?
Describe your task, we will make a research and respond to you as soon as possible.