Search for: All records

Many individuals who are in need of mobility assistance do not have access to the proper wheelchair for their type of mobility disability. There is growing research towards creating smart wheelchairs using a variety of methods, such as biopotential signals or eye tracking for input and LiDAR, ultrasonic sensors, or using a camera to create a map or track position. There have been other methods, such as voice control, sip and puff, and hand gestures, but there are disadvantages of these that can limit their usefulness. Smart wheelchairs should account for collisions, but also emphasize the safety and comfort of the user. In this paper, we review and classify state-of-the-art research in smart wheelchairs. Many machine learning models are used for various parts of wheelchairs, from mapping and signal processing to input classification. Smart wheelchairs rely on various hardware devices, such as eye trackers, electrode caps, EMG armbands, RPLidar, RGB-cameras, and ultrasonic sensors. Some hybrid models use a combination of methods to account for some of their limitations. Some research has leaned towards training games to help teach users. Future work should include improvement of classification methods for various input signals and improvement on the accessibility of the technology. 

This survey paper aims to provide an overview of the current state of EEG (Electroencephalography) signal technology as it relates to people with disabilities. It will highlight the various methods and techniques employed, discussing their advantages and disadvantages. The paper will also examine the applications of EEG technology in assisting individuals with disabilities, specifically focusing on Brain-Computer Interfaces (BCIs) and assistive device control. By understanding the current state of EEG signal technology, we can identify the opportunities and challenges involved in utilizing this technology to improve the lives of people with disabilities 

Neural networks for computer vision extract uninterpretable features despite achieving high accuracy on benchmarks. In contrast, humans can explain their predictions using succinct and intuitive descriptions. To incorporate explainability into neural networks, we train a vision model whose feature representations are text. We show that such a model can effectively classify ImageNet images, and we discuss the challenges we encountered when training it.