Vehicle Counting using Deep Learning Models: A Comparative Study

Abdullah, Azizi; Oothariasamy, Jaison · 2020 · Crossref

DOI: 10.14569/ijacsa.2020.0110784

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 accurate vehicle counting in urban traffic scenarios, where manual methods are resource-intensive and prone to error, while standard deep learning models often suffer from low accuracy due to inconsistencies between pre-training datasets and real-world conditions. The authors aim to identify the most effective deep learning architecture for vehicle detection and improve its performance through transfer learning and data annotation. Specifically, the study compares three popular Convolutional Neural Network (CNN) models—Faster R-CNN, Single Shot MultiBox Detector (SSD), and You Only Look Once (YOLOv3)—and proposes a simple tracking algorithm using Euclidean distance and a virtual reference line to count vehicles crossing a specific threshold. The methodology involves two main experiments conducted on video footage captured in Kuala Lumpur, Malaysia, under varying illumination conditions (morning, day, and night). In the first experiment, the authors evaluated the baseline performance of the three pre-trained models (initialized on the COCO dataset) on 10 video clips. In the second experiment, they applied transfer learning by retraining the best-performing model using a custom-annotated dataset comprising 2,550 images of cars, motorcycles, buses, and trucks. The counting system tracks vehicle trajectories by calculating the minimum Euclidean distance between bounding box centers in consecutive frames and increments the count when a vehicle crosses a defined horizontal reference line. Performance was measured using Vehicle Counting Accuracy (VCA). The results indicate that YOLOv3 was the superior model among the three, achieving an average counting accuracy of 66.29% in the baseline test, significantly outperforming Faster R-CNN (38.12%) and SSD (14.53%). However, the baseline YOLOv3 performance was highly sensitive to lighting, with accuracy dropping sharply during early morning and night conditions due to overfitting on the original training data. After retraining YOLOv3 with the annotated dataset, the system’s average vehicle counting accuracy improved significantly to 80.90%. The study also noted that while SSD offered the fastest processing time (0.135 seconds per frame), its detection accuracy was insufficient for reliable counting. The significance of this work lies in demonstrating that transfer learning with custom-annotated data is essential for adapting pre-trained deep learning models to specific real-world environments, particularly those with challenging illumination. The study concludes that while YOLOv3 is the most effective architecture for this application, its performance is heavily dependent on the relevance of the training data to the target environment. The proposed simple tracking and counting method provides a viable, low-complexity solution for automated traffic monitoring, offering a substantial improvement over using off-the-shelf models without fine-tuning.

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.

• distraction detection algorithms