Airy YOLOv5 for Disabled Sign Detection

Designated parking spaces for individuals with disabilities are only meant to be used by vehicles with proper handicapped signage. Real-time monitoring is necessary to ensure that only authorized vehicles are parked in these spaces and to prevent unauthorized vehicles from using them. First, this research proposes to replace the backbone of a baseline YOLOv5 model which has 9 blocks with 6 EfficientNet blocks with less parameters but still have a higher accuracy in detecting disabled signs among other signages on the windshield of cars. Second, to compensate for the loss of blocks we have included an attention mechanism before detection part in our architecture which allows us to focus on the important regions needed for the task. Additionally, we propose to use a better optimizer AdamW to prevent overfitting. Based on these improvements, we have created a new object detector named Airy YOLOv5. To evaluate the effectiveness of our proposed method, a dataset containing images of cars with disabled signage on their windshields will be gathered and labeled. Experiments using this dataset show that our model achieves a better F1 score of 0.67 with 5 percent less parameters compared to the baseline model.