Optimized Fast R-CNN for Automated Parking Space Detection: Evaluating Efficiency with MiniFasterRCNN

Abstract

Automated parking space detection is a crucial application of computer vision in intelligent transportation systems. In this study, we developed a Fast R-CNN-based model for classifying and localizing parking spaces into empty and occupied categories. The model architecture consists of a pre-trained CNN backbone (ResNet50) for feature extraction, a Region Proposal Network (RPN) for generating potential bounding boxes, and Region-of-Interest (RoI) pooling for feature refinement. The classification head utilizes a softmax activation function with cross-entropy loss, while the bounding box coordinates are refined using smooth L1 loss. To facilitate training, we employed Roboflow for dataset annotation, creating ground truth bounding boxes for parking spaces. The model was fine-tuned using transfer learning, leveraging knowledge from the COCO dataset. Training involved hyperparameter optimization, including learning rate scheduling and weight decay, to enhance convergence. Model selection was based on validation loss and accuracy to ensure generalization to unseen data. The model was deployed using Gradio, allowing real-time parking space detection from uploaded images. Despite achieving a final loss of 0.8280, the model exhibited some background noise distortions, impacting detection accuracy. To address this limitation, we explored a lightweight alternative, MiniFasterRCNN, optimized for efficiency with a simpler architecture. The MiniFasterRCNN was trained on a three-class dataset (empty, occupied, background), achieving a validation accuracy of 77.78%. However, attempts to achieve 100% accuracy proved inefficient, highlighting the need for further improvements, such as segmentation techniques (Masked R-CNN). This research demonstrates the feasibility of Fast R-CNN-based models for parking space detection while emphasizing the importance of architectural optimizations and hyperparameter tuning for improved accuracy and robustness in real-world applications