TL;DR
This paper tackles the critical challenge of accurately detecting and localizing parking spaces in complex, real-world environments for autonomous parking systems. They introduce a new dataset, PSEX, augmented using GANs for diverse conditions, and propose an improved PP-Yoloe model (SimSppf_mepre-Yoloe) that achieves better accuracy and maintains real-time performance on edge devices.
What Problem Does This Paper Address?
Autonomous parking systems rely heavily on precise parking space detection and localization. This is a fundamental prerequisite for enabling fully autonomous parking, but it remains a significant challenge, especially in varied and difficult environmental conditions (like fog, snow, or rain). Existing public datasets often lack this diversity, and current detection methods struggle with accuracy and real-time performance, particularly for small or occluded parking space markers. The paper aims to bridge this gap by creating a more robust dataset and a more capable detection model.
Methods
The core of the proposed solution involves two main components:
- Dataset Augmentation with GANs: To address the limitations of existing datasets, the authors construct a new dataset called PSEX. They leverage Generative Adversarial Networks (GANs), specifically an algorithm called LapStyle, to fuse various extreme environmental conditions (fog, snow, rain, sandstorms) into existing parking images. This style transfer technique aims to enhance image semantics and environmental style, creating a more diverse and realistic dataset.
- Improved PP-Yoloe Model: The authors adapt and enhance the PP-Yoloe object detection model for real-time parking space detection. Key modifications include:
- Replacing the "ResSpp" module with a "ResSimSppf" module in the backbone.
- Introducing a "SimSppf" structure in the neck of the network.
- Replacing the SiLU activation function with ReLU in the CBS (Conv-BN-SiLU) blocks, creating a "CBR" module.
- Adding an auxiliary detection head to the prediction head. These changes are designed to improve the model's ability to detect small targets and handle varying scales.
Datasets
- PSEX (Parking Slot Extended): A newly constructed dataset created by augmenting existing public datasets (like PS2.0) using GAN-based style transfer. It aims to incorporate diverse environmental conditions (fog, snow, rain, sandstorms) and complex scenarios.
- PS2.0: A publicly available parking space dataset from Tongji University, used as a base for augmentation.
- PSV (Panoramic Surround View): Another public dataset from Tongji University.
- PIL_PARK: A vehicle location group dataset used for pre-training the model.
Key Insights and Ideas
- GANs are powerful for environmental augmentation: The paper demonstrates that GANs can effectively fuse diverse environmental conditions into existing images, creating a more robust and realistic dataset for training autonomous driving perception models. This goes beyond simple image transformations.
- Targeted architectural improvements matter: By strategically modifying specific modules (SimSPPF, CBR) and adding an auxiliary head to PP-Yoloe, the authors achieve significant performance gains without drastically increasing computational complexity.
- Pre-training on diverse datasets accelerates convergence: Using the PIL_PARK dataset for pre-training the proposed model led to faster convergence during training on the PSEX dataset.
- Balancing accuracy and speed is crucial for edge deployment: The proposed model strikes a good balance between detection accuracy and real-time inference speed, making it suitable for deployment on resource-constrained edge devices like those found in intelligent vehicles.
Key Results
- Improved Accuracy: The proposed SimSppf_mepre-Yoloe model achieved an average improvement of 4.5% in mAP50 and 2.95% in mAP50:95 compared to the baseline PP-Yoloe across various parking space detection tasks.
- Specific Gains: Notable improvements were seen in detecting parking space occupancy (e.g., parallel parking occupancy increased by 8.33% in mAP50) and corner types (T_corner improved by 9.17% in mAP50).
- Real-time Performance: The model maintains comparable inference latency to the baseline PP-Yoloe. On the Jetson AGX Xavier platform with TensorRT optimization, it achieved up to 33.7 FPS.
- Robustness: The model demonstrated improved performance in complex environments, including heavy snow conditions, where other tested algorithms struggled.
Limitations and Future Work
- Model Quantization: The authors suggest that future work will involve using model quantization techniques to further reduce the model's size and improve its detection speed for deployment on specific edge devices like UAVs.
- Real-time Performance vs. Accuracy Trade-off: While the proposed model offers a good balance, the Raux-Yoloe model achieved higher FPS but with a slight accuracy degradation. Future work might explore further optimizing this trade-off.
- Dataset Diversity: While PSEX is more diverse, continuous efforts to expand and refine datasets with even more challenging scenarios would be beneficial.
Summary updated