Search for a command to run...
Fog remains a major challenge for visual perception systems in autonomous vehicles, surveillance, and environmental monitoring. This paper introduces a hybrid deep learning framework that synergistically combines Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks for robust fog detection and classification. This approach leverages RGB image sequences. Such work may also be applied for fog removal in video, especially in transportation systems on highways to avoid accidents. This work has been carried out through multi-modal inputs through three specialized branches: i) a ResNet-50 backbone for spatial feature extraction, ii) Bidirectional LSTM layers for temporal pattern analysis, and iii) a fog feature encoder incorporating physics-inspired descriptors (Dark Channel Prior, grayscale variance, edge attenuation, and contrast energy) to identify the fog. An augmented dataset of 28,397 samples is used, covering four fog types (clear, homogeneous, inhomogeneous, and sky fog) synthesized from five benchmark sources. Experimental results focus on the state-of-the-art performance, with the CNN-LSTM hybrid achieving better test accuracy on the validation dataset. It also experimented using the Vision Transformer (ViT)-LSTM and dual-CNN model and shows considerable accuracy. During execution, this work is experimented through 37 FPS on NVIDIA GTX 1650, making it suitable for edge deployment. This study also shows a comparative study of the considered models.. This work addresses critical gaps in environmental perception for autonomous systems operating in critical weather conditions to record the fog.