Traffic Accident Image Classification using CLIP with a Zero-Shot Approach

Abstrak

Accident image detection is a critical step in supporting rapid response systems during emergency situations. This study employs the Contrastive Language–Image Pre-training (CLIP) model to detect accident images using a zero-shot approach, without requiring retraining on a specific dataset. The CLIP model leverages multimodal embeddings from text and images, enabling detection based on textual descriptions. The experimental results, using the ViT base patch 32 model, show that this method achieves a Top-1 accuracy of 32% and a Top-5 accuracy of 95.86%. Although the Top-1 accuracy indicates that further optimization is needed, the high Top-5 accuracy demonstrates the significant potential of CLIP for efficient accident image detection. With further development, this technology can serve as a reliable solution for various emergency response scenarios, offering flexibility and efficiency in detecting accident-related images.

Identifier

• DOI: https://doi.org/10.55749/ijdrr.v1i1.185
• OpenAlex: https://openalex.org/W7169762691