Scholarlink

The accurate classification of driving conditions is crucial for the safe and efficient operation of semi-autonomous vehicles. Even for manually operated vehicles, classification of driving conditions is necessary for improving visibility. Variations in conditions, such as heavy rainfall, dense fog, or darkness, present distinct challenges that directly impact driving dynamics and safety. This initiated research on computer vision techniques to classify driving conditions. However, the complexity of road conditions with different traffic levels, sceneries, etc., poses significant challenges for the classification task. In this paper, we propose the use of ConvNeXT to address this problem. We perform transfer learning by using feature extractions from pre-trained ConvNeXT and train the classification module for categorizing the road images into one of four classes, viz. sunny, rainy, foggy, dark. We create training, validation and test sets from the available data from different databases. Due to limited available data, we also employ data augmentations by randomly selected crop regions, random horizontal flips, Gaussian blur with random kernels and standard deviations, and the addition of salt and pepper noise. The performance of ConvNeXT is compared with VGG-16, ResNet-50, DenseNet-121, Inception-v3, Mobilenet-v3, Efficientnet-v2, Xception, and Inception Resnet-v2. ConvNeXT provides the best accuracy of 95.25%, precision of 0.95, and a recall of 0.96, among other competing methods.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Classification of Driving Condition Images for Semi-autonomous Vehicles Through Data Augmentation and Transfer Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ravindra Gautam,
  • Anirban Dasgupta,
  • Ayush Datta Jaiswal,
  • Devansh Sharma

摘要

The accurate classification of driving conditions is crucial for the safe and efficient operation of semi-autonomous vehicles. Even for manually operated vehicles, classification of driving conditions is necessary for improving visibility. Variations in conditions, such as heavy rainfall, dense fog, or darkness, present distinct challenges that directly impact driving dynamics and safety. This initiated research on computer vision techniques to classify driving conditions. However, the complexity of road conditions with different traffic levels, sceneries, etc., poses significant challenges for the classification task. In this paper, we propose the use of ConvNeXT to address this problem. We perform transfer learning by using feature extractions from pre-trained ConvNeXT and train the classification module for categorizing the road images into one of four classes, viz. sunny, rainy, foggy, dark. We create training, validation and test sets from the available data from different databases. Due to limited available data, we also employ data augmentations by randomly selected crop regions, random horizontal flips, Gaussian blur with random kernels and standard deviations, and the addition of salt and pepper noise. The performance of ConvNeXT is compared with VGG-16, ResNet-50, DenseNet-121, Inception-v3, Mobilenet-v3, Efficientnet-v2, Xception, and Inception Resnet-v2. ConvNeXT provides the best accuracy of 95.25%, precision of 0.95, and a recall of 0.96, among other competing methods.