A Dataset for Temporal Semantic Segmentation Dedicated to Smart Mobility of Wheelchairs on Sidewalks

Area of Science:

• Computer Vision
• Robotics
• Smart Mobility

Background:

• Semantic segmentation is crucial for environmental understanding in smart mobility, particularly for autonomous vehicles.
• Existing datasets primarily focus on road-based autonomous vehicles, leaving a gap for other systems like electric wheelchairs (EW).
• The viewpoint significantly impacts semantic segmentation accuracy, especially in diverse mobility environments.

Purpose of the Study:

• To address the lack of specific datasets for electric wheelchairs (EW) on sidewalks.
• To propose a novel dataset and an adapted convolutional neural network (CNN) for improved semantic segmentation.
• To demonstrate the influence of viewpoint on semantic segmentation performance for EW navigation.

Main Methods:

• Development of a new dataset featuring street scenes from sidewalk-level viewpoints in a 3D virtual environment (CARLA).
• Creation of a specialized convolutional neural network (CNN) incorporating temporal processing and accuracy-enhancing techniques.
• Inclusion of a subset with paired images from road and sidewalk viewpoints to analyze viewpoint impact.

Main Results:

• The proposed dataset provides essential data for training and evaluating semantic segmentation models for EW.
• The adapted CNN shows potential for improved accuracy in temporal semantic segmentation tasks.
• Comparative analysis using paired images confirms the significant effect of viewpoint on segmentation outcomes.

Conclusions:

• The developed dataset and CNN are valuable contributions to advancing semantic segmentation for electric wheelchairs (EW) in smart mobility.
• Future research can leverage this dataset to enhance the safety and efficiency of autonomous mobility systems.
• Understanding viewpoint variations is critical for robust environmental perception in autonomous systems.