[Paper Review] Applying Deep Learning to Detect Traffic Accidents in Real Time Using Spatiotemporal Sequential Data

Amir Bahador Parsa, Rishabh Chauhan+3 more|arXiv (Cornell University)|Dec 15, 2019

Traffic Prediction and Management Techniques42 references20 citations

TL;DR

This study applies Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to real-time traffic accident detection using spatiotemporal data from Chicago’s expressways. With an imbalanced dataset of 241 accidents and 6,038 non-accident cases, SMOTE augmentation improved model performance, achieving an AUC of 0.85 for both models, with GRU showing a slight edge in detection rate.

ABSTRACT

Accident detection is a vital part of traffic safety. Many road users suffer from traffic accidents, as well as their consequences such as delay, congestion, air pollution, and so on. In this study, we utilize two advanced deep learning techniques, Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs), to detect traffic accidents in Chicago. These two techniques are selected because they are known to perform well with sequential data (i.e., time series). The full dataset consists of 241 accident and 6,038 non-accident cases selected from Chicago expressway, and it includes traffic spatiotemporal data, weather condition data, and congestion status data. Moreover, because the dataset is imbalanced (i.e., the dataset contains many more non-accident cases than accident cases), Synthetic Minority Over-sampling Technique (SMOTE) is employed. Overall, the two models perform significantly well, both with an Area Under Curve (AUC) of 0.85. Nonetheless, the GRU model is observed to perform slightly better than LSTM model with respect to detection rate. The performance of both models is similar in terms of false alarm rate.

Motivation & Objective

• To develop a real-time traffic accident detection system using deep learning on spatiotemporal traffic data.
• To address class imbalance in traffic accident datasets using SMOTE for improved model generalization.
• To compare the performance of LSTM and GRU architectures in detecting traffic accidents using sequential traffic data.
• To evaluate model effectiveness using AUC, detection rate, and false alarm rate metrics.
• To provide a scalable, data-driven solution for enhancing urban traffic safety systems.

Proposed method

• The study uses LSTM and GRU recurrent neural networks to model sequential traffic patterns from spatiotemporal data.
• Input features include traffic flow, speed, congestion status, and weather conditions collected from Chicago expressways.
• The dataset is highly imbalanced, with 241 accident cases and 6,038 non-accident cases, prompting the use of SMOTE for minority class oversampling.
• Models are trained to classify sequences of traffic data as either accident or non-accident events.
• Performance is evaluated using AUC, detection rate, and false alarm rate to assess discriminative power and reliability.
• The temporal nature of the data is preserved through sequence modeling, enabling real-time inference on streaming traffic data.

Experimental results

Research questions

• RQ1 Can LSTM and GRU models effectively detect traffic accidents in real time using spatiotemporal traffic data?
• RQ2 How does class imbalance in traffic accident data affect model performance, and can SMOTE mitigate this issue?
• RQ3 Which recurrent architecture—LSTM or GRU—yields better detection performance in terms of detection rate and false alarm rate?
• RQ4 To what extent do weather and congestion status improve accident detection accuracy when integrated with traffic flow and speed data?
• RQ5 Can deep learning models achieve high AUC while maintaining low false alarm rates in real-world traffic monitoring?

Key findings

• Both LSTM and GRU models achieved an Area Under Curve (AUC) of 0.85, indicating strong discriminative performance in accident detection.
• The GRU model demonstrated a slightly higher detection rate compared to the LSTM model, suggesting better sensitivity to accident events.
• False alarm rates were similar between the two models, indicating comparable reliability in minimizing false positives.
• SMOTE effectively addressed class imbalance, improving model learning on rare accident cases without overfitting.
• The integration of spatiotemporal traffic, weather, and congestion data enhanced model performance beyond using traffic flow and speed alone.
• The results confirm that recurrent networks are effective for real-time accident detection using sequential traffic data streams.