Deep Learning Techniques for Visual Simultaneous Localization and Mapping Optimization in Autonomous Robots

Abstract

In the previous chapter, we explored the application of reinforcement learning to autonomous robots, focusing on the indoor maps constructed using the Simultaneous Localization and Mapping (SLAM) technique. Visual SLAM (VSLAM) is highlighted as a cost-effective SLAM system that leverages 3D vision to execute location and mapping functions without limitations on distance detection range. VSLAM can also incorporate inertial measurement unit (IMU) measurements to enhance the accuracy of the device's pose estimation, particularly in scenarios where visual data alone is insufficient, such as during rapid movements or temporary visual obstructions. This chapter shifts the focus to integrating deep learning (DL) with VSLAM to boost its accuracy and performance. DL can significantly enhance VSLAM by providing semantic understanding, object detection, and loop closure detection, improving the system's overall situational awareness. We delve into six DL models that are pivotal in advancing VSLAM capabilities: Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Neural Networks (NNs), Graph Convolutional Networks (GCNs), Message Passing Neural Networks (MPNNs), and Graph Isomorphism Networks (GINs). Each of these models offers unique advantages for VSLAM. CNNs are adept at processing visual information and extracting spatial features, while LSTMs excel in handling temporal dependencies, making them suitable for dynamic environments. NNs provide a flexible framework for various learning tasks, and GCNs effectively capture spatial relationships in graph-structured data. MPNNs and GINs enhance the ability to process and analyze complex graph-based data, improving the robot's understanding of its environment. This chapter provides a comprehensive overview of how these DL models can be integrated with VSLAM to achieve more robust and efficient autonomous navigation. Through detailed explanations and practical examples, we illustrate the potential of combining DL with VSLAM to advance the field of autonomous robotics.