Developing and testing deep learning algorithms for object detection and classification on Indian roads comes with unique challenges:

1) Live testing of the algorithm in a dense traffic situations.
2) The camera position was not fixed.
3) No camera calibration was done.
4) Presence of region-specific vehicles such as auto-rickshaws.
5) Implementation on a very simple edge device.

To address this, we extended a pre-trained model by training it with approximately 1200 auto-rickshaw images, enabling it to recognise this special class along with standard vehicles.

The results were promising:
1) The model accurately detected both regular vehicles and region-specific ones like auto-rickshaws.
2) It identified a person even when my face was partially covered with a helmet for a short duration.
3) It detected my bike correctly, even when it appeared in the camera’s view for just a brief moment.

This is one of the AI-powered features we customised as part of our ADAS (Advanced Driver Assistance Systems) solutions.