Vehicle-Type Detection Based on Compressed Sensing and Deep Learning in Vehicular Networks

Yinghua Li, Bin Song, Xu Kang, Xiaojiang Du, Mohsen Guizani2018 · DOI

TL;DR

This paper introduces a novel vehicle-type detection system that combines Compressed Sensing (CS) for efficient saliency map generation with a deep learning Convolutional Neural Network (CNN) for classification. By leveraging CS in the measurement domain, the method significantly speeds up the window calibration process and achieves higher accuracy in vehicle detection, particularly in scenarios with limited training data.

What Problem Does This Paper Address?

The increasing number of vehicles on the road presents significant challenges for traffic regulation and monitoring. Accurate and efficient vehicle detection and classification are crucial for intelligent transportation systems, road safety, and urban sensing applications. Traditional methods often struggle with high computational complexity due to sliding-window approaches and pixel-domain feature extraction, leading to redundant computations and slow processing. Existing deep learning methods, while powerful, can also be computationally intensive and require large amounts of training data. This paper aims to address these limitations by proposing a method that is both computationally efficient and accurate, even with restricted training data.

Methods

The proposed method integrates two key techniques:

  • Compressed Sensing (CS) for Saliency Map Generation: Instead of traditional pixel-domain saliency map generation, which is computationally expensive, this paper utilizes CS theory. CS allows for signal acquisition and recovery from fewer measurements than traditional methods. The authors apply CS to measure the image in a "measurement domain," which is significantly smaller than the pixel domain. This enables the generation of saliency maps at a lower computational cost and faster speed. The saliency map highlights salient regions (potential vehicles) by identifying areas with low correlation within the measurement domain. This approach effectively reduces redundant areas in the image, leading to more efficient target detection.
  • Convolutional Neural Network (CNN) for Classification: Once the salient regions are identified using the CS-based saliency map, a CNN is employed for vehicle-type classification. The paper specifically mentions using the ResNet50 model, known for its ability to handle deeper networks and mitigate training errors. The CNN takes the identified regions and classifies them into different vehicle types. The integration of CS for saliency map generation is intended to speed up the window calibration stages of the CNN-based classification.

Datasets

The paper mentions using the following datasets:

  • Laboratory's intelligent traffic big data project: For training the model.
  • Vehicles from the Internet: Also used for training.
  • MIT CBCL Database: Used for evaluation.
  • Caltech Database: Used for evaluation.

The paper does not introduce any new benchmarks but utilizes existing public databases for comparison.

Key Insights and Ideas

  • Saliency in the Measurement Domain: The core idea is that saliency can be effectively detected in the compressed "measurement domain" generated by CS, rather than solely in the pixel domain. This domain offers significant compression and dimensionality reduction, leading to faster processing.
  • Reduced Redundancy: By using CS-generated saliency maps, the method effectively minimizes redundant areas in the image, making the subsequent window calibration and target detection steps more efficient.
  • Efficiency with Limited Data: The paper suggests that their method reduces requirements for high computation and large amounts of training data, making it practical for real-world applications where data might be scarce or computational resources are limited.
  • Linear Relationship between Domains: The authors theoretically establish an approximately linear relationship between the cross-covariance matrices in the measurement and frequency domains, which underpins the feasibility of analyzing image features in the measurement domain.

Key Results

  • Speedup: The proposed method is shown to speed up the window-calibrating stages of CNN-based image classification.
  • Accuracy: The method demonstrates better overall performance in vehicle-type detection compared to other methods.
  • On the MIT CBCL database, the proposed method achieved an accuracy of 0.9412, outperforming Haar+Cascade (0.9338), raAdaBoost (0.9355), and CS-CNN (0.9371).
  • On the Caltech Database, the proposed method achieved an accuracy of 0.9504, surpassing Haar+Cascade (0.9238), raAdaBoost (0.9302), and CS-CNN (0.9427).
  • Qualitative Results: Figure 6 and Figure 7 show successful saliency map generation and window calibration results, visually demonstrating the method's ability to identify salient regions. Figure 8 showcases effective vehicle-type detection on real traffic monitoring images.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while CS is an excellent dimension reduction tool, the saliency map generation still involves some processing.
  • The method's performance with extremely complex backgrounds or very small images might be influenced by block size selection, requiring careful tuning.
  • Future Work: The paper highlights the broad prospects for practical applications in vehicular networks, implying that further development and deployment in real-world scenarios are envisioned. The authors suggest that the method's efficiency and accuracy, especially with restricted training data, make it a strong candidate for future research and implementation in intelligent transportation systems.

Summary updated

BibTeX
@misc{li2018vehicle,
  author = {Yinghua Li and Bin Song and Xu Kang and Xiaojiang Du and Mohsen Guizani},
  title = {{Vehicle-Type Detection Based on Compressed Sensing and Deep Learning in Vehicular Networks}},
  year = {2018},
  doi = {10.3390/s18124500},
}
Figure