Incident detection using data from social media

Abstract

Due to the rapid growth of population in the last 20 years, an increased number of instances of heavy recurrent traffic congestion has been observed in cities around the world. This rise in traffic has led to greater numbers of traffic incidents and subsequent growth of non-recurrent congestion. Existing incident detection techniques are limited to the use of sensors in the transportation network. In this paper, we analyze the potential of Twitter for supporting real-time incident detection in the United Kingdom (UK). We present a methodology for retrieving, processing, and classifying public tweets by combining Natural Language Processing (NLP) techniques with a Support Vector Machine algorithm (SVM) for text classification. Our approach can detect traffic related tweets with an accuracy of 88.27%.

Exploring the Potential of Twitter to Understand Traffic Events and Their Locations in Greater Mumbai, India

“…Most classifiers for detecting traffic events share some common (and basic) tasks, e.g., data collection, data preprocessing, feature generation, model development. Tweets related to traffic can be collected randomly from public users within a given spatial extent [7], [12], [15] or by using a relevant keyword search [11], [24], or by simply following specific official accounts [18], [33]. Tweets collected using a spatial extent are geotagged whereas tweets collected by keywords or by following some specific user accounts are typically not geotagged.…”

“…D'Andrea and others used inverse document frequency (IDF) as features [11]. Similarly, other researchers used a single word tokens (unigram) and multiple word tokens (n-gram) and their associated term frequencies (TF) as feature vectors [7], [8], [12], [14], [15], [22], [24]. Gu and colleagues used only a selected number of unigrams pertaining to traffic incidents in the US [24].…”

“…They showed that an SVM-based classifier could achieve 95.75% and 88.89% accuracy for two class and three class classification respectively through an n-fold cross validation. Similarly Salas and colleagues developed a single layered tweet classification model using a SVM to categorize tweets into traffic or non-traffic [12]. To train the model, [12] used a balanced data set consists of 871 traffic related tweets and 871 non-traffic related tweets.…”

“…Similarly Salas and colleagues developed a single layered tweet classification model using a SVM to categorize tweets into traffic or non-traffic [12]. To train the model, [12] used a balanced data set consists of 871 traffic related tweets and 871 non-traffic related tweets. While testing, they used 290 traffic and 290 non-traffic related tweets.…”

“…In contrast to [31], we aim to extract all the traffic locations from the relevant tweets while dealing with the informal nature of location mentions in the tweet content. Our work goes beyond [11], [12], [15] and not only detects traffic relevant tweets but also geolocates traffic events using relevant tweet content. To geolocate traffic events, we developed a hybrid georeferencing model, partly motivated by [41], but trained and tested in the context of traffic events in Greater Mumbai in India.…”