A Vehicular Queue Length Measurement System in Real-Time Based on SSD Network

Okaishi, Wahban Al; Zaarane, Abdelmoghit; Slimani, Ibtissam; Atouf, Issam; Benrabh, Mohamed · 2021 · Crossref

DOI: 10.2478/ttj-2021-0003

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 accurately measuring vehicular queue lengths at intersections in real-time, a critical parameter for detecting traffic congestion caused by signals, accidents, or infrastructure issues. The authors motivate their work by highlighting the limitations of traditional traffic detectors, such as inductive loops, which require pavement cutting and have small detection zones, as well as older vision-based systems that rely on traditional image processing techniques like edge or corner detection. These traditional methods are often susceptible to environmental factors such as weather conditions, shadows, and road markings. To overcome these drawbacks, the study proposes a robust, vision-based system that utilizes deep learning for vehicle detection. The proposed system operates in two main steps using video feeds from stationary cameras. First, it employs a frame differencing method to detect motion in specific target areas at the head of each lane. If motion is detected, the system assumes no queue exists. If no motion is detected, the system proceeds to the second step: vehicle detection using a modified Single Shot Multibox Detector (SSD) algorithm. To enhance performance for this specific application, the authors modified the standard SSD architecture. They adjusted the aspect ratios to 1.5 and 2 to better fit vehicle shapes, removed the last two feature maps to focus on smaller objects typical in queue scenarios, and redesigned the base network to improve the detection of small vehicles. The system calculates queue length by summing the heights of the bounding boxes of detected vehicles, extending backward from the head of the queue until no vehicles are found. Experimental results demonstrate the superiority of the proposed method over existing techniques. The vehicle detection module was trained and tested on the MIO-TCD dataset, with 12,000 images for training and 8,000 for testing. Comparative analysis against Faster R-CNN, YOLO2, and standard SSD512 showed that the modified SSD achieved the highest performance metrics: a Recall Rate of 79.6%, a Precision Rate of 87.89%, and an F1-score of 0.84, while also requiring the least processing time per frame (0.2 seconds). Furthermore, the system’s ability to measure queue length was evaluated using video data from traffic cameras in Casablanca. When compared to a system using edge detection, the proposed SSD-based system provided measurements that more closely followed the ground-truth queue length, demonstrating greater accuracy and stability in real-time conditions. The significance of this work lies in its contribution to intelligent transportation systems by providing a reliable, real-time method for queue length estimation that is robust against environmental variations. By integrating motion detection with a tailored deep learning model, the system offers higher accuracy and efficiency than traditional image processing approaches. This capability supports better traffic management and signal control, addressing the need for precise, quantitative analysis of traffic scenes without the physical constraints of embedded sensors.

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.

• traffic density