Overview / Usage

The death toll in road accidents every year is more then that in any other man made or natural disaster taken cumilatively. People tend to avoid helping injured people in road accdient due to the fact that the might then be called by police for interogation or might get stuck into troublesome situations if the victim dies. A report says that in most of the cases the cause why doctors loose the patient is because they are not brought to the hospitals on time. My application uses camera feeds on traffic lights and lanes to detect when an accident occurs and instantaniously informs the responsible person reguarding the location time and images of casualties.

Methodology / Approach

I handpicked images from videos for a reliable dataset. Trained that dataset using pretrained weights of resnet18 , 34 ,51 and 101 an also vgg16. I took the performance measures of all after fine tuning and epoch stablisation. Then i simulated the process of an accident in a desktop app developed in kivy to take images every secomd and detect if an accident occured or not. I recieved an atmost accuracy of 77 percent with a dataset of 450 images in all.

Technologies Used

fastai was used for model training and testing. Kivy was used for video feed sampling and prediction. The training was done on a 12gb tesla gpu on google colab and some on kaggle. Django for admin site for notifications and images of accidents along with place and camera number

Repository

https://github.com/mridularya/road-accident-detection-.git