Deep Learning for Marine Vehicles Parking Availability: A ResNet50-Based Deep Feature Engineering Model

Deep Learning for Marine Vehicles Parking Availability: A ResNet50-Based Deep Feature Engineering Model

In this research, our essential objective is to evaluate the availability of parking spaces within port/marine/fisher shelters employing a novel computer vision-based approach. Therefore, we collected a new dataset and developed a ResNet50-based image classification model to detect parking status. Initially, we collected a new image dataset using an unmanned aerial vehicle (UAV) from over 200 fisher shelters and the collected image dataset contains two classes which are parking available or not (full). To automatically detect parking available fisher shelters, a new ResNet50-based deep feature engineering (DFE) approach has been recommended. In the recommended DFE approach, we introduced a novel semi-overlapped patch division strategy to extract local features like transformers. To implement this model, we first trained the ResNet50 approach on our collected training dataset and a trained ResNet50 model has been obtained. Subsequently, deep features have been derived using the proposed semi-overlapped patch division approach and the global average pooling (GAP) layer of the trained ResNet50. Nine feature vectors have been generated using patches and a feature vectors has been extracted from the whole image. By using this strategy, we have generated both local and global features and these features have been merged to create the ultimate feature vector. To select informative features from the generated ultimate feature vector, iterative neighborhood component analysis (INCA) feature selector has been applied. The chosen features by INCA were employed as input of the support vector machine (SVM), is a shallow classifier, classifier to create classification results. The used ResNet50 convolutional neural network (CNN) attained 100% training accuracy and 94.23% validation accuracy. Subsequently, the recommended DFE model was assessed on test images, achieving a test classification accuracy of 97.27%. Furthermore, we utilized Grad-CAM and feature analysis to provide interpretable results for the presented model. The achieved classification performance and the explanatory outcomes demonstrably illustrate the capability of the proposed model for automatic detection of parking availability in fisher shelters. These findings support the utility of computer vision as a viable solution for this application.