This project built a system that can detect any road accident occurrence in CCTV camera videos. It will reduce the delay in providing medical aid to the victims of road accidents by informing the relevant authorities whenever it detects an accident.
In a report of the World Health Organization (WHO), Approximately 1.35 million fatalities occur during road accidents. 93% of the fatality counts belong to low and middle-income countries. This is because of the lack of awareness in people living in these countries on what to do when a road accident occures. This lack of awareness of the bystanders causes a delay in response to the victim, which results in the loss of lives.
If we could reduce this delay by detecting accidents automatically and sending the information of the accident to the traffic authorities or the nearest hospital, countless lives could be saved.
- Compile a data set from different sources (local vehicle data as well as coco data).
- Annotate the data set in required format.
- Use GANs to create Synthetic Night data.
- Train YOLOv4 to perform obect detection on vehicles(including local vehicles) and people.
- Perform object detection on CADP data set.
- Train a Conv-LSTM model to detect accidents on the results of object detection.
- Create a pipeline to perform inference, Object Detection -> Accident Detection on CCTV Videos.
Please refer to Accident_Detection Notebook to know more about the Accident Detection approach.
Working with Accident data from South-East Asian Countries can be quiet challenging due to the types of vehicles and the congestion of traffic. Sometimes, a frame can contain tens and hundreds of objects. Detecting objects and then accidents from such videos can be very difficult with limited resources like us.
There is a lot of room for improvements. Steps to further improve the performance can be found at the last section.
MAP (Mean Avg. Precision) and Loss on Unseen Data:
Confusion Matrix:
Epoch-Accuracy:
🟠 Train 🔵 Validation
Epoch-Loss:
🟠 Train 🔵 Validation
- Tensorflow 2.0
- Open CV
- darknet
- Numpy
- Setup
darknetby following the guide present here. - Replace the
yolo-obj.cfg, present in the "cfg and weights" directory, in darknet. - Replace the Object Detection weights available here
- Install all the dependencies.
- Run Prediction.py with path to input and output files
- Segment out vehicles and nearby objects and walls and feed those to LSTM.
- Compile even more data for that purpose, the more the merrier.
- Another approach is to work on Accident Localization instead of Object Detection and localization.