Robust Autonomous Driving Control using Auto-Encoder and End-to-End Deep Learning under Rainy Conditions

A key element of autonomous vehicles is lane-keeping. To keep an automobile in its lane, numerous artificial intelligence technologies and vision systems are employed. Despite being widely used, current systems function effectively in favorable weather. Downstream computer vision models may perform worse due to rain and other weather-related obstacles. For this reason, rain, snow, and mist noise must be detected, recognized, and removed by vision systems to improve system performance. For these issues, the autoencoder neural network is particularly common. The main goal of using an autoencoder for image deraining is to maintain originality because autoencoders adhere to the backpropagation procedure instead of the traditional techniques, which results in the signal coming out being ultimately the altered version of the input signal. This study covers a method for reducing the amount of rain in photos utilizing image processing, deep learning methods with autoencoder implementation, and end-to-end learning to determine the best steering rotation setting for retaining a vehicle in its lane using a donkey car. After training, the end-to-end model may be able to directly guide the car using the information from the front view camera, unlike the conventional method which manually breaks down the autonomous driving issue into specific details. The convolutional neural network (CNN) model takes raw images as input and generates the proper steering angles. The AI Kit Donkey Car is used in this study to collect data, train models, and evaluate them. According to the experimental results, the combination of the deraining method and end-to-end deep learning produces good performance in rainy situations.