# Collision Detection: An Improved Deep Learning Approach Using SENet and ResNext

**Authors:** Aloukik Aditya, Liudu Zhou, Hrishika Vachhani, Dhivya Chandrasekaran, Vijay Mago  
**arXiv:** [2201.04766](https://arxiv.org/abs/2201.04766)  
**Date:** 2022-01-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.04766v1/

## Key points

- Vehicle collisions are one of the leading causes of death worldwide due to increased population and traffic on roadways.
- The automotive industry is developing collision detection and prevention systems using sensors and computer vision advancements.
- Aloukik Aditya, Liudu Zhou, Hrishika Vachhani, Dhivya Chandrasekaran, and Vijay Mago propose an improved deep learning approach for collision detection.
- Their model comprises ResNext architecture with SENet blocks and outperforms popular deep learning models like VGG16, VGG19, Resnet50, and stand-alone ResNext.
- The proposed model achieves a ROC-AUC of 0.91 using significantly less proportion of the GTACrash synthetic data for training compared to existing baseline models.
- This reduces computational overhead while maintaining high accuracy.
- The authors cite previous research that uses data from various sources to efficiently predict accidents and provide assistance to drivers.
- They trained their deep learning-based approach using synthetic data and compared its performance against other models using real-world datasets like Cityscapes dataset for semantic urban scene understanding.
- Overall, the proposed model shows promise in improving collision detection systems by leveraging advancements in deep learning techniques.

## Summary

The increased population and traffic on roadways have made vehicle collisions one of the leading causes of death worldwide. To address this issue, the automotive industry is developing techniques to build collision detection and collision prevention systems using sensors and advancements in computer vision. In a recent article, Aloukik Aditya, Liudu Zhou, Hrishika Vachhani, Dhivya Chandrasekaran, and Vijay Mago propose an improved deep learning approach for collision detection. Their model comprises ResNext architecture with SENet blocks and outperforms popular deep learning models like VGG16, VGG19, Resnet50, and stand-alone ResNext. The proposed model achieves a ROC-AUC (receiver operating characteristic - area under curve) of 0.91 using significantly less proportion of the GTACrash synthetic data for training compared to existing baseline models. This reduces computational overhead while maintaining high accuracy. The authors also cite previous research that uses data from various sources to efficiently predict accidents and provide assistance to drivers. They trained their deep learning-based approach using synthetic data and compared its performance against other models using real-world datasets like Cityscapes dataset for semantic urban scene understanding. Overall, the proposed model shows promise in improving collision detection systems by leveraging advancements in deep learning techniques. Its high accuracy with reduced computational overhead makes it a potential candidate for implementation in real-world scenarios to prevent accidents on roadways.