Fast Multiclass Vehicle Detection on Aerial Images

Liu, Kang; Máttyus, Gellért · 2015 · OpenAlex-citations

DOI: 10.1109/lgrs.2015.2439517

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 detecting vehicles in aerial images for applications such as traffic management and urban planning. The authors identify key difficulties, including the small size of vehicles relative to the image resolution, complex backgrounds in man-made areas, and the need for near real-time processing on large images without auxiliary data like road databases or precise ground sampling distance (GSD) information. The goal is to develop a method that not only detects vehicle bounding boxes but also estimates their orientation and type (car vs. truck) efficiently on standard hardware. The proposed method employs a two-stage pipeline. The first stage utilizes a fast binary sliding window detector based on Integral Channel Features (ICF) within a Soft Cascade structure using an AdaBoost classifier. To handle arbitrary vehicle orientations, the authors compare a single classifier trained on all directions against an aggregated classifier approach using multiple detectors trained on specific orientations. The second stage refines the detected bounding boxes by estimating orientation using Histogram of Oriented Gradients (HOG) features and a neural network, followed by a multiclass classifier to distinguish between cars and trucks. The system operates on original, non-orthorectified frame images, requiring no prior georeferencing or scale information. Experiments were conducted on a dataset of 20 aerial images over Munich, Germany, captured by a DLR 3K camera system, as well as a UAV dataset. The Munich images had a resolution of 5616 × 3744 pixels with an approximate GSD of 13 cm. Results show that the aggregated classifier method with eight detectors (covering 180-degree rotation steps) yielded the best detection performance, achieving a recall of 69.3% and precision of 86.8% on the Munich dataset, significantly outperforming a Viola-Jones baseline. The method demonstrated robustness across different image scales and achieved high accuracy in orientation estimation, with the most common errors occurring in opposite-direction classifications. For type classification, the system achieved 95.1% accuracy on the Munich dataset. Crucially, the method processed a 21-megapixel image in approximately 4.4 seconds on a single-threaded laptop, which is substantially faster than compared methods that required GPUs or longer processing times. The significance of this work lies in its ability to provide comprehensive vehicle information—location, orientation, and type—rapidly and automatically from raw aerial imagery without external dependencies. This capability supports real-time traffic monitoring and infrastructure management. The authors conclude that the integration of ICF in a Soft Cascade structure offers an optimal balance of speed and accuracy, suggesting that future improvements could involve deep neural networks for further refinement.

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