[Paper Review] Automated Approach for Computer Vision-based Vehicle Movement Classification at Traffic Intersections

Udita Jana, Jyoti Prakash Das Karmakar+5 more|arXiv (Cornell University)|Nov 17, 2021

Video Surveillance and Tracking Methods4 citations

TL;DR

This paper proposes an automated, unsupervised computer vision approach for classifying vehicle trajectories at traffic intersections into movement types—such as right-turn, left-turn, and through movements—using hierarchical clustering and a novel similarity measure. The method eliminates manual region-of-interest specification and demonstrates robust performance across diverse traffic scenarios without reconfiguration.

ABSTRACT

Movement specific vehicle classification and counting at traffic intersections is a crucial component for various traffic management activities. In this context, with recent advancements in computer-vision based techniques, cameras have emerged as a reliable data source for extracting vehicular trajectories from traffic scenes. However, classifying these trajectories by movement type is quite challenging as characteristics of motion trajectories obtained this way vary depending on camera calibrations. Although some existing methods have addressed such classification tasks with decent accuracies, the performance of these methods significantly relied on manual specification of several regions of interest. In this study, we proposed an automated classification method for movement specific classification (such as right-turn, left-turn and through movements) of vision-based vehicle trajectories. Our classification framework identifies different movement patterns observed in a traffic scene using an unsupervised hierarchical clustering technique Thereafter a similarity-based assignment strategy is adopted to assign incoming vehicle trajectories to identified movement groups. A new similarity measure was designed to overcome the inherent shortcomings of vision-based trajectories. Experimental results demonstrated the effectiveness of the proposed classification approach and its ability to adapt to different traffic scenarios without any manual intervention.

Motivation & Objective

• To address the challenge of classifying vehicle trajectories by movement type in traffic scenes using vision-based data.
• To eliminate reliance on manually defined regions of interest in existing classification methods.
• To develop a robust, automated framework that adapts to varying traffic conditions and camera calibrations.
• To improve accuracy and generalization in vehicle movement classification without human intervention.

Proposed method

• The framework employs unsupervised hierarchical clustering to identify distinct movement patterns from vehicle trajectories.
• A custom similarity measure is designed to handle inherent variability in vision-based trajectory data from different camera calibrations.
• Trajectories are assigned to movement groups based on a similarity-based assignment strategy using the proposed metric.
• The method is fully automated, requiring no manual labeling or region-of-interest configuration.
• The approach is evaluated across multiple traffic scenarios to test adaptability and robustness.
• The framework is trained and tested on real-world traffic camera data to ensure practical applicability.

Experimental results

Research questions

• RQ1 Can an unsupervised clustering approach effectively identify movement patterns in vision-based vehicle trajectories without manual region-of-interest definition?
• RQ2 How well does the proposed similarity measure handle variations in trajectory data due to differing camera calibrations?
• RQ3 To what extent can the classification framework generalize across diverse traffic intersection scenarios without reconfiguration?
• RQ4 What is the accuracy and robustness of the automated classification system in real-world traffic environments?

Key findings

• The proposed method achieves high classification accuracy across multiple traffic scenarios without requiring manual tuning or region-of-interest specification.
• The novel similarity measure effectively captures movement patterns despite variations in trajectory data caused by camera calibration differences.
• The unsupervised clustering approach successfully identifies distinct movement types—such as right-turn, left-turn, and through movements—based on trajectory shapes.
• The framework demonstrates strong adaptability to different traffic conditions, maintaining consistent performance across diverse intersection layouts.
• Experimental results confirm the method’s robustness and scalability in real-world deployment settings.
• The system operates fully automatically, reducing dependency on human expertise for setup and configuration.