The problem Anti-Vehicle Accident Detection using Gestures solves

The issue of driver drowsiness, distraction, and inattentiveness during driving poses a significant risk to road safety. According to the National Highway Authority of India (NHAI), driver fatigue and distraction are major factors in approximately 20% of all fatal car crashes in India. The need is to monitor driver behavior and alert them when signs of fatigue or distraction are detected. Additionally, the model should also suggest actions such as rest breaks or changes in lighting or music to prevent further distraction which would improve road safety and save lives.

Challenges we ran into

The following challenges were faced:

1. Dataset Collection: To address and resolve such huge social-issue, we needed proper and well elaborated dataset collection that includes areas of diverse situations. However, it was very hard for us to locate such idealistic dataset. Fortunately, after due research, we found MRL Eye Dataset to cater our needs and train our model with its collection of diverse dataset.

2. Data Pre-Processing: This is a very time taking process because the amount of data for training was very limited. Thus, we had concerns with underfitting and to deal with such issue, we used the concept of Data Augumentation through which we were able to generate multiple images from single image by changing values of Hyper-Parameters such as Zoom, rotation, etc.

Since, it is emperically proved that neural network converges better for normalised input values so we used the concept of feature scaling to make conversions efficient.

3. Model Selection: We had a basket of models during training of datasets which had to be choosen. Earlier, we used harcascad face recognition model to train our dataset but it was not providing satisfactory accuracy. Then our mentor reccommended us to use MTCNN, which worked pretty well and gave us ~95 percent accuracy overall.