Novel On-Road Vehicle Detection System Using Multi-Stage Convolutional Neural Network

Jisu Kim, Sungjun Hong, Euntai KimIEEE Access 2021 · DOI

Abstract

Modern deep learning-based object detection systems often require significant GPU resources, making them difficult to deploy in cost-sensitive commercial automotive environments. This paper introduces the Multi-Stage Convolutional Neural Network (MSCNN), a lightweight detection pipeline designed for real-time operation on standard CPUs. By utilizing a four-stage process that exploits vehicle geometric properties, MSCNN achieves high accuracy, particularly for small, distant objects, reaching 72.1% Average Precision (AP) on the KITTI benchmark.

Introduction

While models like YOLOv3 and Faster R-CNN have set benchmarks for accuracy, their reliance on high-end GPUs poses challenges for commercial integration due to power, heat, and cost constraints. Furthermore, these models often struggle with small object detection; as images pass through deep layers, pooling operations reduce resolution, causing distant vehicles to lose critical features.

The MSCNN addresses these issues by replacing a single deep network with a cascade of specialized shallow networks. This approach preserves high-resolution features and incorporates geometric priors to focus computation on relevant image regions.

Methodology: The Multi-Stage Architecture

The MSCNN framework evolves from traditional detection pipelines by integrating CNN features directly into the region proposal and classification stages.

Comparison of object detection architectures, highlighting the evolution toward integrated CNN-based systems.

The system operates in four distinct stages:

  1. Vehicle Lower-Boundary Detection: A shallow CNN scans vertical image strips to predict the bottom edge of vehicles.
  2. Vehicle Upper-Boundary Detection: A second CNN predicts the top edge. This is optimized by the Position-Wise Vehicle Height Predictor (PVHP), which uses Bayesian statistics to estimate vehicle height based on vertical position in the frame.
  3. Region Proposal Network (RPN): Specific anchor boxes are generated based on the detected boundaries, significantly reducing the search space compared to generic sliding windows.
  4. Vehicle Classification: A final CNN classifies these proposals, filtering out false positives.
Visual breakdown of the four-stage detection process, from boundary estimation to final classification.

Key Insights

  • Geometric Optimization: By recognizing that vehicles follow predictable perspective rules (e.g., distant cars appear higher in the frame), the PVHP eliminates redundant processing in areas like the sky.
  • Resolution Preservation: Shallow networks maintain spatial detail, which is crucial for detecting vehicles with widths $\le$ 50 pixels.
  • CPU-Efficiency: The authors prioritize the "AP-FPS" metric, ensuring that the model remains performant on standard hardware like an Intel i5 CPU.

Results and Evaluation

The MSCNN was evaluated on the KITTI VD, CrowdAI, and AUTTI datasets. Key findings include:

  • Performance: Achieved 72.1% AP on KITTI.
  • Speed: Maintained ~13.5 FPS (74ms/frame) on an Intel i5-4670 CPU, outperforming YOLOv3-tiny in precision while matching its speed on non-GPU hardware.
  • Small Object Detection: Significantly outperformed deeper models in detecting distant vehicles where resolution loss typically occurs.

Conclusion

The MSCNN provides a viable path for real-time vehicle detection on low-power hardware. While the sequential nature of the stages introduces a dependency on the initial boundary detection, the system's ability to leverage geometric constraints offers a significant advantage in efficiency. Future work will focus on temporal tracking and expanding the model to detect pedestrians and cyclists.

Summary updated

BibTeX
@article{kim2021novel,
  author = {Jisu Kim and Sungjun Hong and Euntai Kim},
  title = {{Novel On-Road Vehicle Detection System Using Multi-Stage Convolutional Neural Network}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3093698},
}
Figure