K-Fold Vehicle Collision Prediction – ResNet

A deep learning model leveraging ResNet to predict vehicle collisions from dashcam footage with high accuracy

Project Description

This project involves real-time vehicle collision predictions based on the BDD100K dashboard camera footage. A ResNet50-based CNN model was employed and trained based on the extracted frames from the video and k-fold cross-validation.

To achieve optimal results, the use of Optuna-based hyperparameter optimization and manually-designed data augmentation approaches, such as random blurring, brightness adjustments, and spatial cropping, was incorporated to mimic real-world conditions.

The resulting model had an accuracy of 81.15% on the test set and minimized the false positives by 18%. The performance of the model was evaluated on the basis of the precision-recall curves, confusion matrix, and dashboard analysis using Matplotlib and Seaborn.