CNN-based Visual Localization for Autonomous Vehicles under Different Weather Conditions

Abstract

• A Convolutional Neural Network (CNN) was developed for autonomous localization layer-by-layer in urban driving situations. • To check the positional accuracy of the CNN, RGB images are combined with depth images using the IHS method. • With an accuracy rate of 94.74%, the simulation results demonstrated the effectiveness of the suggested strategy. • The Simulation findings demonstrate the superiority of the suggested technique for different weather conditions. Autonomous vehicles (AV) are expected to improve, transform, and revolutionize ground transportation. Previous techniques are dependent on localization employing pricey inaccuracy Global Positioning sensor. Furthermore, the performance loss is caused by drifting errors of Simultaneous Localization and Mapping. Regarding categorizing and analyzing texture, cameras are much more accessible and practical. This work contributes to obtaining high accuracy for AV localization and reducing errors in predicted positions. Based on the light, accurate, and robust proposed Convolutional Neural Network (CNN), it will scale down the computational complexity and shorten the training time. Considering various weather and time of day conditions such as bright sunny, hard rain noon, and wet cloudy noon with a vision-only system Red, Green, and Blue (RGB) low-cost camera sensors. To check the positional accuracy of the CNN, RGB images are combined with depth images using the IHS method. The k-Mean technique evaluates the similarity between a specific image and all street images to obtain precise coordinates. The Simulation findings demonstrate the superiority of the suggested technique for different weather conditions, which has an accuracy of up to 94.74% and a Mean Squared Error MSE in a distance of 0 meters, as opposed to [19], where the MSE in the projected position is 4.8 meters. Another indication of the proposed method's effectiveness is that it yielded reliable results when its validity was tested on images from a dataset that had not been trained.