Real-time critical moving object identification and dynamic path planning for assisting visually impaired people
File(s)
Author(s)
Surougi, Hadeel
Type
Thesis
Abstract
Autonomous objects like cars, motorcycles, bicycles, and pedestrians pose risks to visually impaired people (VIPs). Existing camera-based assistive solutions fail to identify fast-moving threats, namely Critical Moving Objects (CMOs), and cannot plan optimal collision-free local paths, essential for guiding VIPs around CMOs.
This thesis presents AR-AwareAvoid, a novel AI-based assistance scheme to assist VIPs in safely navigating outdoor environments, including different classes of CMOs moving at varying speeds. AR-AwareAvoid comprises two components: ARAware, which effectively identifies CMOs and gives the VIP more time to avoid danger through simultaneously addressing real-time CMO identification, CMO risk level classification, and a risk-aware prioritised CMO warning notification, and MinD, which plans the shortest path required to avoid CMOs while considering the VIP’s mobility constraints, CMO types and movement patterns, and predicted collision times. MinD also conducts a look-ahead safety prediction of the optimal path to ensure more safety for VIPs.
Real-world experiments demonstrate that ARAware accurately identifies CMOs (97.26% mAR and 88.20% mAP), precisely classifies CMOs according to their risk levels (100% mAR and 91.69% mAP), timely warns about high-risk CMOs while effectively reducing false alarms by postponing low-risk CMOs’ notifications, and operates in real-time (only takes 31ms processing time on an NVIDIA RTX 2080 GPU. Compared to the DEEP-SEE, ARAware achieves significantly higher CMO identification accuracy (by 42.62% in mAR and 10.88% in mAP), with 93% faster end-to-end processing speed. Additionally, simulation results demonstrate that MinD outperforms the Artificial Potential Field (APF) approach in effectively planning a short collision-free route, requiring only 2.75m of movement on average, shorter than APF by 84.47%, with a 0% collision rate, adapting to the VIP’s mobility limitations, and provides a high safe time separation (≥ 4.80s on average compared to APF). MinD also shows real-time performance, with decisions taking only 1.20ms processing time on the same GPU.
This thesis presents AR-AwareAvoid, a novel AI-based assistance scheme to assist VIPs in safely navigating outdoor environments, including different classes of CMOs moving at varying speeds. AR-AwareAvoid comprises two components: ARAware, which effectively identifies CMOs and gives the VIP more time to avoid danger through simultaneously addressing real-time CMO identification, CMO risk level classification, and a risk-aware prioritised CMO warning notification, and MinD, which plans the shortest path required to avoid CMOs while considering the VIP’s mobility constraints, CMO types and movement patterns, and predicted collision times. MinD also conducts a look-ahead safety prediction of the optimal path to ensure more safety for VIPs.
Real-world experiments demonstrate that ARAware accurately identifies CMOs (97.26% mAR and 88.20% mAP), precisely classifies CMOs according to their risk levels (100% mAR and 91.69% mAP), timely warns about high-risk CMOs while effectively reducing false alarms by postponing low-risk CMOs’ notifications, and operates in real-time (only takes 31ms processing time on an NVIDIA RTX 2080 GPU. Compared to the DEEP-SEE, ARAware achieves significantly higher CMO identification accuracy (by 42.62% in mAR and 10.88% in mAP), with 93% faster end-to-end processing speed. Additionally, simulation results demonstrate that MinD outperforms the Artificial Potential Field (APF) approach in effectively planning a short collision-free route, requiring only 2.75m of movement on average, shorter than APF by 84.47%, with a 0% collision rate, adapting to the VIP’s mobility limitations, and provides a high safe time separation (≥ 4.80s on average compared to APF). MinD also shows real-time performance, with decisions taking only 1.20ms processing time on the same GPU.
Version
Open Access
Date Issued
2024-04-25
Date Awarded
01/10/2024
License URL
Advisor
McCann, Julie
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)