|
--- |
|
license: apache-2.0 |
|
--- |
|
|
|
# Traffic Accident Detection |
|
## Overview |
|
The [DETR](https://huggingface.co/facebook/detr-resnet-50) (DEtection Transfomer) model utilized in this implementation serves as a sophisticated solution for accident detection. This state-of-the-art model leverages the power of transformers, originally designed for natural language processing, to excel in object detection tasks. Trained on a diverse dataset, the DETR model demonstrates its capability to identify and locate objects within images, particularly excelling in the crucial task of accident detection within traffic scenes. |
|
|
|
Harnessing advanced computer vision techniques, DETR offers unparalleled accuracy and efficiency in recognizing potential incidents, providing valuable insights for enhancing road safety. Its utilization is pivotal in real-time monitoring and analysis, empowering applications geared towards automated accident detection and response systems. |
|
|
|
This DETR model is equipped with a robust post-processing pipeline, incorporating Non-Maximum Suppression (NMS) to refine detections and deliver precise and actionable results. Combined with efficient inference times, this DETR model stands as a powerful tool in the realm of accident detection, contributing to the development of intelligent and safety-focused systems in various domains. |
|
|
|
## Dataset |
|
Introducing a cutting-edge approach to accident detection, this model employs the DETR (DEtection Transfomer) architecture, specifically designed to seamlessly identify accidents within a comprehensive scene captured in a single image. Unlike traditional methods, this innovative model operates within the context of full images, leveraging the power of transformer-based object detection. |
|
|
|
Table 1: When we use dataset focuses on accident label, model fails to detect accidents when traffic jams. |
|
| traffic jams | traffic jams | |
|
|-------|-------| |
|
| ![traffic jams](./demo/traffic-jams-3.png) | ![traffic jams](./demo/traffic-jams-4.png) | |
|
|
|
Table 2: When we use multi label dataset (accident and vehicle), model can detect accidents accurately without reducing detection performance when traffic jams |
|
| traffic jams | traffic jams | accident | accident | |
|
|-------|-------|------|-------| |
|
| ![traffic jams](./demo/traffic-jams-1.png) | ![traffic jams](./demo/traffic-jams-2.png) | ![accident](./demo/accident-1.png) | ![accident](./demo/accident-2.png) | |
|
|
|
Trained on a diverse and multilabel dataset, including 'accident' and 'vehicle' labels, the model excels in simultaneously recognizing both accident-related incidents and the presence of vehicles. This dual-label dataset enhances the model's capacity to comprehensively understand and interpret complex traffic scenarios, making it a potent tool for real-time accident detection and analysis. |
|
|
|
By adopting a holistic perspective on the entire image, this DETR-based model contributes to a more robust and nuanced understanding of potential accidents, fostering advancements in automated safety systems. Its proficiency in detecting accidents within the broader context of traffic scenes positions it as a valuable asset for applications dedicated to enhancing road safety and emergency response. |
|
|
|
[![try our dataset](https://img.shields.io/badge/roboflow%20traffic%20accident%20dataset-download-purple?logo=hackthebox)](https://universe.roboflow.com/hilmantm/traffic-accident-detection) |
|
[![try it online](https://img.shields.io/badge/huggingface%20spaces-try%20it%20online-blue?logo=tryitonline)](https://huggingface.co/spaces/hilmantm/detr-traffic-accident-detection) |