# Real time road scene classification and enhancement for driver assistance under adverse weather

## TL;DR

This paper introduces a low-cost system that uses machine learning to classify and enhance road scenes in real time, improving visibility for drivers in adverse weather conditions.

## Contribution

A novel, high-accuracy road scene classification and enhancement system using the Random Committee algorithm on affordable hardware.

## Key findings

- The proposed system achieves 98.67% classification accuracy using the Random Committee algorithm.
- YOLO-based object detection performs better on enhanced images compared to original ones.
- The system is implemented on a low-cost Raspberry Pi 5 with a USB camera and 7-inch display.

## Abstract

Highways are the most widely used mode of transportation worldwide, accounting for the majority of passenger movement. However, the drivers often face difficulties due to poor visibility through the windshield under adverse conditions. In such situations, an alternative mode of vision is essential, and a video display showing the roadway is ideal for this purpose. This paper presents an efficient machine learning-based classification system for various road scenarios, including daytime, nighttime, foggy, and rainy conditions. After classifying the scenario, enhancement techniques are applied to improve the visibility of the road image, ensuring clarity in all atmospheric conditions. Various machine learning algorithms were tested for accuracy in classifying road scenarios, and the most accurate one was selected. Following classification, specific image enhancement techniques were applied to improve the road video according to the identified scenario. A high-intensity mapping technique was used for glare reduction, and a low-light enhancement technique was applied for better night visibility. Defogging and deraining algorithms were employed for foggy and rainy conditions, respectively. An affordable, low-cost system was developed based on the Raspberry Pi 5, utilising a USB camera and a 7-inch display. Compared to state-of-the-art techniques such as Resnet-101 and custom CNN applied for the same kind of work, the proposed model achieves a classification accuracy of 98.67% using the Random Committee algorithm, demonstrating superior performance in roadway classification, even on limited hardware. This approach also shows strong potential for integration into ADAS systems, especially in autonomous vehicles, where larger image datasets and more generalised machine learning or deep learning-based enhancement techniques can be applied. The improved performance of YOLO-based object detection on enhanced images, compared to the original ones, further validates the effectiveness of this method.