Edge Computing-Enabled FHWA Vehicle Classification on Surveillance Videos

Abstract: Traffic surveillance videos are useful in many transportation applications. To address the absence of techniques capable of accurately classifying all 13 Federal Highway Administration (FHWA) vehicle categories in videos, the authors at West Virginia University have developed a two-stage vision-based method for precise vehicle classification, leveraging both semantic and geometric features extracted from surveillance videos. Initially, semantic features, such as vehicle appearance, are utilized to categorize vehicles into six general types using a deep learning model—these include motorcycle, passenger car, van, pickup, bus, and truck. Subsequently, geometric features, focusing on axle configuration, are employed to further differentiate trucks into specific types aligned with the FHWA classification scheme. To support this, a dictionary encompassing all conceivable truck axle distribution patterns is established, enabling accurate determination of fine-grained truck classes through dictionary searches. The study also incorporates specialized filtration of lift axles to account for their impact on vehicle classification. Additionally, a verification scheme is introduced to validate vehicle classification results, filtering out misclassified vehicles and enhancing overall classification accuracy. Field experiments were conducted to assess the proposed method's performance, revealing an impressive overall accuracy of 98.6% in classifying all 13 FHWA vehicle categories, with an additional 1.1% improvement facilitated by the verification scheme.

Nevertheless, the current method operates on a cloud computing platform, which, while effective, presents challenges for real-time deployment due to latency and reliance on high-bandwidth connections. To overcome these limitations, this study explored the transition of the data processing pipeline to an edge computing framework. In this study, we systematically profiled the performance of the current cloud-based implementation to establish baseline metrics, identify computational bottlenecks, and assess potential areas for efficiency enhancement. Optimization strategies were then applied, focusing on minimizing memory and processing demands through model compression, selective data processing, and algorithmic refinement. These strategies were tested iteratively to maximize performance under the constraints of edge devices. Results demonstrated the feasibility of implementing FHWA-compliant vehicle classification directly on edge devices, achieving substantial gains in processing speed and system responsiveness while maintaining high accuracy. This work offers a crucial step toward scalable, real-time traffic monitoring solutions for a variety of transportation environments.