Traffic Accident Detection and Prediction using Deep Learning Techniques

Abstract

Road accidents have increased by a prominent number around the globe from the past few years and have become a major reason for the harm and loss of human life. To encounter this problem, the researchers are trying to practice the advanced technologies like artificial intelligence and machine learning for the detection and prediction of road accidents. In this digital era, social media is being used excessively and it has produced an enormous corpus of geo-tagged data. This paves way for the developers and practitioners to research on the detection and prediction of accidents owing to availability of the data on social media like Facebook, Twitter etc. The social media posts can be actively used for research related to accidents as they often comprehend pertinent and useful information about accidents. Modern research has proposed the detection of traffic accidents by the use of partial feature set that can be prolonged using emotions, sentiments, and additional numerous extended approaches in order to attain higher accuracy in deep learning techniques. This research aims to use deep learning models for the detection and prediction of road accidents and achievement of overall accuracy of results by using extended features like time information, weather condition, geo coded location of accidents etc. The proposed scheme enhances the accuracy and efficiency by making use of advanced techniques like LSTM, CNN, and GRU layers. For prediction of accidents, 75% accuracy has been achieved. Moreover, addition of extended features increased the accuracy of deep learning model for accident detection by 8%, making test accuracy to reach at 94%. Our early investigations also shows that Twitter can be used as a medium to predict Traffic accidents. Our early investigations also show that Twitter has the ability to provide accident-related content. To summarize, incorporating social network data into a traffic-related analysis opens up a plethora of new avenues for transportation science. The findings indicate that social network data may be noisy due to some limitations like usage of slangs, sentence structure, limited characters in a post etc. but if they are integrated with modern techniques of Deep Learning like LSTM, CNN etc. They can prove to be a beneficial in detection and prediction of traffic accidents. The model developed in this study can be used to detect traffic accidents in real time, possibly leading to faster emergency responses and decision making. Much more precise models will be calibrated in the future by making dataset that can be widely utilized for study.