A novel approach for end-to-end navigation for real mobile robots using a deep hybrid model

The ability to use a single camera for indoor navigation of a real mobile robot can help develop a complete navigation system, eliminating the need to use a traditional navigation system. The latter involves many subsystems such as perception for capturing the scene and extracting the necessary information, path planning to determine the optimal trajectory from the origin to the destination, and obstacle avoidance. The complexity of this system makes it highly sensitive to variations in the environment, making it inappropriate for real-time use, particularly for mobile robots with limited computational capacity. In this paper, we present an end-to-end navigation system for indoor navigation based on convolutional neural networks (CNNs). We developed four distinct deep hybrid model (DHMs) approaches using four pretrained convolutional networks (Visual Geometry Group 19 (VGG19), Residual Network 50 (ResNet50), Alex Krizhevsky’s network (AlexNet), and Efficient Network B0 (EfficientNetB0)) as feature extractors, combined with a random forest (RF) machine learning classifier. Validation metrics (accuracy, precision, and F1 score) were used to evaluate the performance of the models. During the test phase, we introduced three new and never-before-seen environments. This study revealed that the combination of VGG19 and random forest models produced excellent results, achieving an accuracy of 97.63%. This model was also able to navigate efficiently in most environments, generating a smoother path than other models.