Temporal Collage Prompting: A Cost-Effective Simulator-Based Driving Accident Video Recognition With GPT-4o

This paper presents temporal collage prompting, a novel approach for detecting and classifying simulator-based driving accident videos using GPT-4o. While recognizing accident videos is crucial for assessing drivers’ abilities and safety, this task traditionally relies on human labor. In addition, it is time-consuming and inefficient, especially when dealing with numerous videos. Large multi-modal models (LMMs) offer a promising solution to reduce processing time but face a challenge with context window limitation when handling video data. We address this by developing a method that optimizes input efficiency while preserving temporal information. Our approach combines multiple video frames into a collage, significantly reducing input tokens. Testing with custom scenarios generated from CARLA, a driving simulator, we achieve 93% accuracy in accident recognition using a 2x2 collage at 1 frame per second (FPS), outperforming a uniform frames baseline method. This configuration reduces token usage by 91% compared to uniform frames at 3 FPS, while improving accuracy from 72% to 93%. Through ablation studies with various collage layouts and sampling rates, we found that lower frame rates, particularly 1 FPS, are more effective for this task. Our results demonstrate that optimized frame sampling and collage creation can enhance both efficiency and accuracy of video recognition using LMMs, offering a promising solution for simulator-based driving video recognition in driving assessment, with potential applications in other domains requiring temporal visual recognition. We provide our data and source code for public use¹.