Optimizing Car Collision Detection Using Large Dashcam-Based Datasets: A Comparative Study of Pre-Trained Models and Hyperparameter Configurations

The automatic identification of traffic collisions is an emerging topic in modern traffic surveillance systems. The increasing number of surveillance cameras at urban intersections connected to traffic surveillance systems has created new opportunities for leveraging computer vision techniques for automatic collision detection. This study investigates the effectiveness of transfer learning utilizing pre-trained deep learning models for collision detection through dashcam images. We evaluated several state-of-the-art (SOTA) image classification models and fine-tuned them using different hyperparameter combinations to test their performance on the car collision detection problem. Our methodology systematically investigates the influence of optimizers, loss functions, schedulers, and learning rates on model generalization. A comprehensive analysis is conducted using 7 performance metrics to assess classification performance. Experiments on a large dashcam-based images dataset show that ResNet50, optimized with AdamW, a learning rate of 0.0001, CosineAnnealingLR scheduler, and Focal Loss, emerged as the top performer, achieving an accuracy of 0.9782, F1-score of 0.9617, and IoU of 0.9262, indicating a strong ability to reduce false negatives.