A real-time autonomous highway accident detection model based on big data processing and computational intelligence

Özbayoğlu, Ahmet Murat; Küçükayan, Yusuf Gökhan; Doğdu, Erdoğan · 2016 · OpenAlex-citations

DOI: 10.1109/bigdata.2016.7840798

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of traffic congestion caused by accidents, proposing a real-time autonomous highway accident detection system. The authors argue that early detection of incidents can save lives, reduce response times, and improve traffic efficiency. To achieve this, they developed a model leveraging big data processing and computational intelligence techniques to predict accident occurrences based on real-time traffic flow data, aiming to provide alerts before official notifications arrive. The study utilizes 2015 traffic flow data from seven Road Traffic Microwave Sensors (RTMS) located on a 15-mile section of the TEM highway in Istanbul. This specific highway segment was chosen because it lacks traffic lights or sharp curves, ensuring that slowdowns are likely due to disruptions. The dataset comprises over 100 million rows of raw data, including vehicle counts, average speeds, and occupancy ratios recorded every two minutes. Due to data inconsistencies and volume, the authors employed an Extract-Transform-Load (ETL) process using HDFS and Apache Spark on a 10-PC cluster to clean and standardize the data. Feature extraction focused on differential values (changes between consecutive readings) for velocity, occupancy, and road capacity, alongside weekday/weekend indicators. These features were used to train three computational intelligence models: nearest neighbor, regression tree, and feed-forward neural networks. The dataset included 72 recorded incidents, resulting in a highly imbalanced classification problem with 276,354 total data points. The results demonstrate that all three models achieved high accuracy, generally exceeding 99%, but struggled with a high rate of false positives. For instance, the nearest neighbor model with equal bias detected 13 of 14 test accidents but generated 13,468 false alarms. Adjusting loss thresholds to reduce false alarms improved precision but decreased recall; a neural network model with a high loss threshold (0.94) reduced false alarms to 580 but missed more than half of the actual accidents. The authors note that false positives may stem from unreported minor issues, weather changes, or sensor malfunctions. Despite the high false alarm rate, the system is deemed useful for status verification, as checking a potential alert is low-cost compared to the benefit of rapid emergency dispatch. The significance of this work lies in demonstrating the feasibility of using big data and machine learning for real-time accident detection. The authors conclude that while the preliminary models are effective, future improvements should include additional features such as meteorological data and road topology, as well as more advanced algorithms like recurrent neural networks. The study highlights the trade-off between recall and precision in imbalanced traffic data and suggests that even imperfect detection systems can provide valuable early warnings for traffic management departments.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• telematics crash prediction