A novel collision avoidance approach for powered wheelchair steering using deep learning

Malik Haddad, David Sanders, Giles Tewkesbury, Martin Clifford Langner, William Malcolm Keeble

Research output: Contribution to journalArticlepeer-review


This paper presents a novel collision avoidance approach for steering a powered wheelchair using a Deep Learning architecture. The architecture consisted of a 6-layer Artificial Neural Network and used a Rule-based method to create training and testing sets. An ultrasonic sensor array was created using three ultrasonic transducers. The transducers measured the distance from the wheelchair to the nearest obstacles to the right, in front, and to the left of the wheelchair. Readings from the ultrasonic array were used as inputs to the Neural Network. The architecture applied Deep Learning to steer a powered wheelchair away from obstacles. Six possible steering directions were considered: forward, turn left, spin left, turn right, spin right and stop. The outcome of the new architecture steered the wheelchair away from obstacles. The new approach behaved satisfactorily when tested against training and testing sets and provided 99.17% training accuracy and 97.53% testing accuracy. Testing showed that the new approach successfully steered a powered wheelchair away from obstacles. A user could over-ride the new system if required. Clinical trials will be conducted at Chailey Heritage Foundation.
Original languageEnglish
Number of pages7
JournalJournal of Physics: Conference Series
Publication statusAccepted for publication - 1 Nov 2021
Event2nd International Symposium on Automation, Information and Computing 2021 - China, Beijing , China
Duration: 3 Dec 20216 Dec 2021


  • UKRI
  • EP/S005927/1


Dive into the research topics of 'A novel collision avoidance approach for powered wheelchair steering using deep learning'. Together they form a unique fingerprint.

Cite this