Human Robot Interactions using Efficient Semantic Mapping

Identificador:  TDX:4383

Handle: https://hdl.handle.net/20.500.11797/TDX4383

Autors:  Singh, Aditya

Resum:
Comparamos los resultados de odometría utilizando el conjunto de datos de Kitti, mientras que los conjuntos de datos de NYU-D y Camvid se utilizan para entrenar el modelo. Para el mapeo semántico de escenas, proponemos una arquitectura basada en Fusion multimodelo que emplea tres columnas vertebrales convolucionales para clasificar escenas con sus etiquetas correspondientes. El LoCobot, un robot asequible y de código abierto disponible en el CIR (Centro de Robótica Inteligente), IIIT-A. Perception-based modeling of the workspace is a crucial requirement for mobile robots to navigate indoor environments. In order to enable robots to effectively interact with humans, it is also necessary to have a semantic description of the environment. This thesis presents affordable semantic mapping techniques for robots, enabling them to interpret the environment and interact meaningfully with it. We explore various approaches for learning semantics, including neural networks based deep learning, and rule-based systems. As an alternative to resource-intensive deep learning models, we propose the use of lightweight deep learning models such as TF-Lite and YOLOv3. These models are integrated into the robotic design and aesthetics to generate object-wise semantic maps, focusing on two-dimensional representations. By predicting bounding boxes and calculating changes in odometry using image sequences captured from the robot's camera, we provide a detailed representation of the scene. To enhance the scene's details and facilitate odometry extraction, we propose an encoder-decoder model that predicts depth and semantic labels per pixel.