Real-time traffic, accident, and potholes detection by deep learning techniques: a modern approach for traffic management

Abstract

The practical applications of social media have raised the bar for real-time event detection all over the globe. It has been deemed useful for extracting important data that enables people across the world to access information within mere seconds of its occurrence. Among the many use cases of the aforementioned applications lies detection of issues regarding the smooth mobility of traffic on the road. Machine learning models that were developed earlier used support vector machines with Bag of Words. However, BoW (Bag of Words) suffers from problems such as the inability to manage semantic relationships between words, limited representation, and high dimensional representation. Our model developed for a similar cause fetches the required data from Twitter to make the target authorities aware of the issues like potholes, accidents, and high traffic density (congestion) of the Chandigarh tri-city area. The data collected then undergoes multiple stages of pre-processing. After that, multiple word embedding models come into play to build semantic relationships between the words and intra software jargon. The resultant is then processed through the several recently introduced deep learning based natural language processing. The current study aims to perform a comparative evaluation of the several advanced state-of-the-art classification models at the target traffic event detection activity. The developed models make a fact-based prediction to make multi-class classification into the aforementioned categories. From the comparative evaluation, it has been observed that the proposed language processing model (RoBERTa based) pipeline outperforms the existing approaches and is 97% accurate at classifying the real-time tweets. Moreover, the proposed pipeline achieves 96% recall at segregating the traffic events efficiently.