Reckless Status Calculation by Using Yolo and Model Trained on RNN

Abstract

Reckless driving significantly contributes to road accidents and fatalities worldwide, including in Nepal. To tackle this critical issue, we developed reckless driver detection application, an innovative system utilizing advanced machine learning techniques, data analysis to detect and analyze reckless driving behaviors by use of statistical parameters such as standard deviation, mean velocity, variance, kurtosis, skewness, peak-to-peak, comparison mean. The electron app employs the YOLO model to gather information for data analysis for accurate vehicle detection. Once vehicles are detected, Kalman filters track their movements to provide reliable data on speed and trajectory. Pandas is used for saving necessary information in .xlsx files. OpenCV is used for processing the video into its analyzed form with boundaries over vehicles and labels over it, feature extraction such as fps, width, height of the frame. The system further analyzes time-series data using recurrent neural networks trained model to generate a "Reckless Status," quantifying reckless behaviors. During evaluation, the custom RNN model achieved a test loss of 0.2550, accuracy of 94.12%, precision of 1.0, recall of 91.6%, and an F1-score of 0.8979. The ROC curve attained a score of 0.93, indicating strong classification performance. These results demonstrate the effectiveness of the model in identifying reckless driving behaviors with high precision and reliability. This comprehensive solution aims to significantly reduce road accidents and protect vulnerable road users in Nepal.