Predicting car collisions based on footage recorded by dashboard cameras:Participating in the Nexar Dashcam Crash Prediction Challenge on Kaggle

Details

Supervisors

Pecheur, Charles

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enPredicting traffic accidents is a longstanding challenge with major implications for road safety and the development of advanced driver-assistance systems (ADAS). With approximately 1.19 million road deaths worldwide each year, the ability to predict imminent collisions from dashboard camera footage represents a crucial step toward safer and more autonomous vehicles. This thesis addresses this problem as part of the Nexar Dashcam Crash Prediction Challenge hosted on the Kaggle platform. The challenge provides a real-world dataset composed of 1500 training videos and 1344 test videos. The objective is to anticipate collisions in dashcam recordings ending 0.5, 1 or 1.5 seconds before the event occurs, with models evaluated on their ability to predict accidents as early as possible. To address this challenge, multiple deep learning architectures are studied, including hybrid CNN-LSTM models, 3D convolutional networks (3D-CNN) and video transformers (ViViT and VideoMAE). The most effective approach is built on the VideoMAE vision transformer, pre-trained in a self-supervised manner on the Kinetics-400 dataset and fine-tuned on 16-frame clips extracted from the competition videos. Performance is measured by mean Average Precision (mAP). The best configuration reaches 0.827 mAP on the public test set and 0.849 on the private test set, ranking 5th out of 237 teams in the final leaderboard.