Feasibility of Synthetic Data Generation With Image to Video Models for Lane Change Driver Intent Prediction

Abstract

The development and validation of vehicle interior and exterior perception systems requires the collection, synthesis, and modeling of large amounts of diverse data. For many years, this has been achieved primarily through real-world data collection or through the generation of realistic 3D graphics. Today, Generative AI enables the creation of new video content from a given prompt. The generation of synthetic data for the training and validation of neural networks has been well investigated, especially for single-frame perception and object detection tasks. This work focuses on generating synthetic video sequences from a single image and a prompt, and a novel evaluation of the feasibility of such data for validating more spatiotemporal demanding algorithms, specifically lane-change driver intent prediction. The evaluation was carried out on two datasets: Brain4Cars from Stanford and Cornell University, and the Advanced Vehicle Technology Naturalistic Driving Study from MIT. Two generative models were used: CogVideo and OpenSora. Results were compared across three levels: human video review, gaze vector statistics, and a state-of-the-art LSTM-based driver intent prediction algorithm. The findings indicate that current image-to-video models offer promising support for data augmentation and supplementary validation, however, their limited preservation of fine-grained spatiotemporal cues prevents them from serving as standalone substitutes for real-world driver-monitoring data.