# Robust Detection, Assocation, and Localization of Vehicle Lights: A Context-Based Cascaded CNN Approach and Evaluations

**Authors:** Akshay Gopalkrishnan, Ross Greer, Maitrayee Keskar, Mohan Trivedi  
**arXiv:** [2307.14571](https://arxiv.org/abs/2307.14571)  
**Date:** 2023-07-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.14571v1/

## Key points

- The paper addresses the importance of vehicle light detection in autonomous driving tasks.
- The authors experiment with different CNN architectures, data augmentation techniques, and contextual preprocessing methods to improve accuracy and reduce confusion with surrounding vehicles.
- The model is evaluated on the LISA Lights dataset, allowing for thorough assessment of performance across various lighting conditions and shapes of vehicle lights.
- An average distance error from the ground truth corner of 5.09 pixels is reported, corresponding to approximately 17.24% of the size of the vehicle light on average.
- The proposed method shows promising results in accurately detecting and localizing vehicle lights.
- The authors suggest integrating this model into a pipeline that includes both vehicle detection and vehicle light center detection.
- This research contributes to enhancing safety in autonomous driving systems by improving understanding and prediction of vehicles' light states during critical maneuvers.

## Summary

The paper titled "Robust Detection, Association, and Localization of Vehicle Lights: A Context-Based Cascaded CNN Approach and Evaluations" addresses the importance of vehicle light detection in autonomous driving tasks. To improve accuracy and reduce confusion with surrounding vehicles, the authors experiment with different convolutional neural network (CNN) architectures, data augmentation techniques, and contextual preprocessing methods. The authors evaluate their model on the LISA Lights dataset which allows them to thoroughly assess its performance across various lighting conditions and shapes of vehicle lights. They report an average distance error from the ground truth corner of 5.09 pixels which corresponds to approximately 17.24% of the size of the vehicle light on average. Overall, the proposed method shows promising results in detecting and localizing vehicle lights accurately. The authors suggest integrating this model into a pipeline that includes both vehicle detection and vehicle light center detection to create a comprehensive network for identifying trajectory-informative signals in driving scenes. This research contributes to enhancing safety in autonomous driving systems by improving the understanding and prediction of vehicles' light states during critical maneuvers like lane changes or turns.