This paper presents a method for classifying vehicular trajectories as normal or abnormal in completely uncontrolled video footage. The methodology integrates the YOLOv8 (You Only Look Once version 8) [1] model for vehicle detection with a novel centroid-based interpolation technique and histogram matching for enhanced vehicle tracking. The detected vehicles are labeled and tracked across frames, even in cases where YOLOv8 fails to detect certain vehicles due to low confidence scores. The generated trajectories are fitted with piecewise curvilinear sub-trajectories, and the coefficients of these fitted curves are fed for the classification of different types of vehicular motion trajectories using Long Short Term Memory (LSTM) based autoencoder, achieving high accuracy in distinguishing normal and abnormal traffic patterns. Experimental results are very promising for 15 trajectory classes with 98.6% (96.28%) accuracy, 99% (97.6%) precision, 98.8% (96.28%) recall, and 0.98 (0.96) F1 score when the entire (only 70%) length of the tracks are utilized. These results demonstrate the robustness and applicability of the proposed method in real-world traffic surveillance.
Traffic Vehicle Tracking and Trajectory Classification Using LSTM
2024-12-19
1954801 byte
Conference paper
Electronic Resource
English
Traffic Accident Prediction Using Vehicle Tracking and Trajectory Analysis
British Library Conference Proceedings | 2003
|LSTM-Based Trajectory Tracking Control for Autonomous Vehicles
SAE Technical Papers | 2022
|