Deep Learning Navigation & Detection for Autonomous Vehicles in Extreme Weather

Abstract

In creating decision-making systems that work for autonomous cars regarding the weather, weather detection systems (WDS) play a salient role in adverse weather conditions. Deep learning techniques specify the modalities of allowing autonomous vehicles to understand and appreciate what happens outside under different weather scenarios. This should provide adaptive decision-making concerning various dynamics in the environment, which is very pivotal for autonomous systems. The framework for detection as deep learning-based proposed in this article improves the accuracy of recognizing weather conditions, especially under adverse conditions. This framework proposes transfer learning for using the computational potential of an Nvidia GPU to assess diverse models via three distinct deep Convolutional Neural Networks, namely SqueezeNet, ResNet-50, and EfficientNet. Further evaluation will be made with the two most recent weather imaging datasets, DAWN2020 and MCWRD2018, which together consist of six weather categories: rain, sand, cloudy, snow, sunny, and sunrise. Our experimental analysis validates the claim that high classification accuracy is ensured for all three models. Interestingly, the ResNet-50 CNN model outruns the others by a wide margin of 98.51% precision, 98.48% accuracy, and 98.41% sensitivity along with an extremely short detection time of just about 5 ms during inference on the GPU. We have provided a systematic discussion regarding the comparison of our proposed model with several other pre-trained models and the accuracy gains accrued from it. Cross weather categories, a range of improvements in classification accuracy, between 0.5% and 21%, is recorded. Altogether thus, it is a practicable and dependable solution fast execution for autonomous vehicles in object detection, which therefore is imperative for decisions in a dynamic and complex environment.