Overview / Usage

The project aims to enhance the vision of the driver with smart algorithms that can work in low light scenarios – dusk, night, in lights from headlights etc. We hope to build a segmentation model that:

1. Can be used in real time situations.

2. Is affordable and can run in low cost hardware machines.

3. Detects objects on the road with a good accuracy.

​

Methodology / Approach

Our current software pipeline is listed below:

1. Apply image enhancement algorithms before training
2. Choose an appropriate deep learning model, implement and then train it on these enhanced images.
3. During inference, apply image enhancement techniques.
4. Inference then involves predicting the input data which might be in the form of an image or a video.

Image enhancement algorithms considered:

• Histogram equalisation
• Contrast Enhancement using Exposure Fusion framework
• Retinex based end to end deep learning model for image enhancement.

The Contrast Enhancement using Exposure Fusion framework provided the best mix of quality and speed between the three and hence, was adopted.

Other links

• Contrast Enhancement Algorithm using Exposure Fusion Framework Paper
• Retinex based end to end deep learning model Paper
• DeepLab-V3+ Paper