Automated Vehicle Recognition with Deep Convolutional Neural Networks

Adu-Gyamfi, Yaw Okyere; Asare, Sampson Kwasi; Sharma, Anuj; Titus, Tienaah · 2017 · Crossref

DOI: 10.3141/2645-13

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 need for accurate, nonintrusive vehicle classification systems to support transportation planning, pavement design, and traffic operations. Traditional intrusive sensors are being replaced by video-based systems, which historically offer only coarse classification (up to four classes). The authors propose a vision system capable of finer classification according to the Federal Highway Administration (FHWA) scheme, decoupling object recognition into localization and classification tasks. The methodology employs a two-stage pipeline. First, object localization is performed using Selective Search to generate class-independent region proposals from video frames. Second, these regions are warped to a fixed size (256 × 256 pixels) and processed by a Deep Convolutional Neural Network (DCNN) to extract 4,096-dimensional feature descriptors. These features are then classified using linear Support Vector Machines (SVMs). The DCNN was pretrained on the ILSVRC2012 dataset and fine-tuned on a domain-specific dataset comprising CCTV footage from Iowa and Virginia. The training data included seven merged vehicle classes (motorcycles, passenger cars/SUVs, pickups/vans, buses, single-unit trucks, single-trailer trucks, and multitrailer trucks) and was collected under varying conditions, including different times of day, weather, traffic densities, and camera angles. Experimental results on a test set of 30 videos demonstrate high performance. The system achieved an average precision of 95% and a recall rate of 93%. Specific class performance varied: motorcycles and buses achieved 100% precision, while passenger cars and SUVs reached 95% precision. Single-unit trucks achieved 96% precision, and single- and multitrailer trucks ranged between 92% and 94% precision. Vans and pickups had the lowest precision at 82%, largely due to visual similarities between covered pickups and SUVs. Receiver operating characteristic analysis indicated that the system performs best under free-flow conditions, during daytime or nighttime, and with good video resolution. The system proved robust to challenging conditions such as stop-and-go traffic and varying lighting, though occlusions and poor resolution contributed to missed detections, particularly for smaller vehicles like motorcycles. The study concludes that deep learning techniques can significantly enhance video-based vehicle classification, achieving accuracy comparable to radar and microwave systems. This approach provides transportation agencies with a cost-effective, nonintrusive method for obtaining detailed vehicle type data required for federal monitoring and infrastructure management. The findings highlight the potential of DCNNs to overcome limitations of traditional image processing methods, such as the need for complex camera calibration and susceptibility to shadows.

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