An Improved Vehicle Classification Method Based on Gabor Features

Yingnan Zhao, Zhengdong Liu, Jingyu YangOpenalex Unknown 2006 · DOI · OpenAlex

Abstract

Vehicle classification is an important issue in the domain of ITS (Intelligent Transportation Systems). In this paper we presents an improved one based on Gabor features, which contains three consecutive stages: vehicle segmentation, Gabor features extraction and template matching. A novel non-even sampling of Gabor features is proposed. The experimental data show that this method can heavily reduce the computation and memory requirements, and illustrate good performance both in discrimination ability and robustness.

Introduction

Vehicle classification is a cornerstone of modern Intelligent Transportation Systems (ITS), essential for toll collection, parking management, and traffic monitoring. While Gabor filters are highly effective at extracting discriminative features—mimicking the way the human primary visual cortex processes textures and edges—they come with a heavy "computational tax." Standard Gabor feature extraction typically samples the entire image uniformly, producing massive, high-dimensional feature vectors. For real-time applications or systems with limited hardware, this leads to slow processing speeds and excessive memory usage. This research proposes a way to maintain the discriminative power of Gabor filters while significantly reducing redundant data.

Methodology

The proposed system follows a three-stage pipeline:

1. Vehicle Segmentation

To handle varying lighting, background subtraction is performed on logarithmic intensities. An "opening" morphological operation is then applied to smooth edges, and the vehicle is clipped from the frame using cumulative histograms.

2. Gabor Feature Extraction

The system utilizes Gabor wavelets at three frequencies and eight orientations to capture multi-scale and multi-directional texture information.

3. Non-even Sampling

The core innovation is a non-even sampling scheme based on edge density. Using a Sobel operator to detect edges, the image is divided into:

  • Key Sampling Windows (KSW): Areas with high edge density, sampled densely to capture structural details.
  • Assistant Sampling Windows (ASW): Areas with fewer edges, sampled sparsely. By focusing computational resources on the vehicle's structural outlines, the algorithm generates a much smaller feature vector.

Experimental Results

The method was evaluated using a dataset of four vehicle types: sedan, truck, minibus, and autobus.

Efficiency and Dimension Reduction

The non-even sampling approach successfully reduced feature dimensionality without a proportional loss in accuracy. By adjusting the sampling intervals (e.g., 4:6, 4:8, 6:8) and the proportion of KSW, the researchers found optimal balances between performance and computational cost.

Performance and dimensionality comparison for various sampling intervals and KSW proportions.

Accuracy and Robustness

In clean conditions, the proposed method achieved a 97.8% recognition rate, outperforming baseline regular sampling methods. The performance gap became even more pronounced under environmental disturbances such as global blur (simulating motion), local glance, and highlighting (simulating glare).

Comparison of recognition rates (%) under different environmental disturbances.

As shown in the table above, under "Highlighting" conditions, the proposed method maintained 80.9% accuracy, whereas the baseline dropped to 71.4%. Similarly, it handled "Global blur" significantly better (73.2% vs. 68.4%).

Conclusion

The proposed non-even sampling of Gabor features offers a significant improvement in efficiency for vehicle classification. By prioritizing information-dense edges, the system reduces memory and compute requirements while enhancing robustness against common environmental challenges in ITS. Future work will focus on automating the selection of key points using training samples to create a more adaptive sampling mask.

Summary updated

Abstract

Vehicle classification is an important issue in the domain of ITS (Intelligent Transportation Systems). In this paper we presents an improved one based on Gabor features, which contains three consecutive stages: vehicle segmentation, Gabor features extraction and template matching. A novel non-even sampling of Gabor features is proposed. The experimental data show that this method can heavily reduce the computation and memory requirements, and illustrate good performance both in discrimination ability and robustness.
BibTeX
@inproceedings{zhao2006improved,
  author = {Yingnan Zhao and Zhengdong Liu and Jingyu Yang},
  title = {{An Improved Vehicle Classification Method Based on Gabor Features}},
  booktitle = {Openalex Unknown},
  year = {2006},
  doi = {10.1007/0-387-23152-8\_61},
}
Figure