Methods of Augmentation and Analysis of Deep Learning Based Stray Animal Recognition Architecture for Mishap Prevention

  • Unique Paper ID: 175376
  • Volume: 11
  • Issue: 11
  • PageNo: 2914-2921
  • Abstract:
  • According to reports from the world bank, 64.13% of India's population resides in rural areas. As a result, domestic and stray animals such as cows and buffaloes are frequently seen on or near roadways. It is crucial for drivers or intelligent vehicular systems to detect these stray animals to regulate speed. Each year, hundreds of people and stray animals are injured or killed in vehicle-animal collisions, both during the day and at night. Records from the veterinary department show a nearly 23% increase in such incidents over the last six years. This paper explores a deep learning-based system for detecting stray animals to alert drivers. Deep learning algorithms generally require large datasets for training, testing, and validation. Given the limited availability of public datasets, various data augmentation techniques could be used. For augmentation, we applied techniques such as horizontal flipping, color space transformation, rotation, shear, zoom, intensity transformation, and resizing. The model's performance was evaluated using the training and validation of both datasets over different epochs and batch sizes. The model achieved a true positive rate of 80% to 85%, with 92.5% accuracy on the non-augmented dataset and 91% accuracy on the augmented dataset when tested on real-world camera footage.

Related Articles