Developing a Real-Time Perception System for Indian Roads
The author built an AI-powered perception system called Orvex to handle the unique challenges of driving in India. The system struggled with common benchmarks and assumptions, highlighting the need for AI models tailored to local driving conditions.
Why it matters
This highlights the need for AI systems to be designed and tested for local driving conditions, rather than relying on generic benchmarks.
Key Points
- 1Orvex is a multi-camera real-time perception system designed for Indian urban driving
- 2The system failed in unexpected ways, like scoring all drivers as 'aggressive' and struggling with non-linear vehicle trajectories
- 3Benchmarks based on Western driving data do not capture the true complexity of Indian road conditions
Details
The author built Orvex, a real-time perception system with multiple camera feeds to handle the chaos of Indian urban driving. During testing, the system encountered several issues that exposed the limitations of existing AI benchmarks and assumptions. The driver behavior scoring module, calibrated on Western norms, labeled all drivers as 'aggressive' due to common maneuvers like hard braking and lane changes. The Time to Collision (TTC) calculations also failed to account for the non-linear, opportunistic trajectories of motorcycles and other vehicles. Additionally, the system had to track over 1,100 distinct objects in a single 50-minute drive, far exceeding the object densities seen in popular benchmarks like nuScenes. The author concludes that AI models developed and tested on Western driving data are not well-suited for the unique challenges of Indian roads, and new benchmarks and approaches are needed.
No comments yet
Be the first to comment