The problem Accident Zone alert and accident detection solves

The Accident alert app can be used by everyone so that they are aware when they have entered an accident prone zone, so that they can be more careful and reduce speed, etc. The app alerts the driver whenever he/she enters an accident prone zone. There is also driving mode which will lock the screen if the driver is in an accident prone zone. This is required because using phone while driving especially in accident prone zone can be dangerous. This feature can be disabled if the user is not driving the car. This app will definitely decrease the accidents as prevention is much better.

The accident detection model will detect whether an accident has occured or not, we can use this model to connect with the cctv cameras in india and whenever an accident occurs the police can be sent an alert. In this project we have made the machine learning model. This model will help save many lives because as soon as an accident occurs, the ambulance/police can reach to the location for help. The job of police will become a lot easier as they will not have to check the cctv footage 24/7. Instead this model can check for accidents and alert them.

Challenges we ran into

The challenges we faced while making the mobile app was that initially the app was not running in the background, so whenever the user destroys the app, it stopped working and this is not useful for an app which is using gps and alerting the driver regulary.
So to overcome this challenge we found a solution called foreground services, this helped us resolve the problem and using this the app started working in backhround even when the app was destroyed. The notification is always shown of the app that it is running.

While making the accident detection app, we faced a lot of problems. We were trying to make the model using video object detection but due to a lot of errors, insufficient data, yolov3 errors we were not able to make the model using video object detection. The training data took more than 2 hrs everytime even though we kept low epochs and most of the time it went unsuccessful, so we tried using another method .We were able to make the model using image classifier, and we are classifying whether the frame of video has accident or not.

Also collecting data was a tough task, because the kind of data we wanted was not available a lot on the internet. So during object detection we had to manually label images using labelIMG.

So for now, the model is ready and for demo purpose we used a video in which for each frame we will get to know whether accident is happening or not. This model can be used in cctv cameras.