Depth-weighted detection of behaviours of risk in people with dementia using cameras

Abstract

The behavioural and psychological symptoms of dementia present significant health and safety risks in residential care. While video cameras enable automated behaviours of risk detection, current systems suffer from frequent false alarms or false positives. This challenge arises because events closer to the camera yield disproportionately high reconstruction errors compared to distant events, creating usability issues that hinder real-world deployment. To address this, the aim of this study was to develop a detection system that minimizes false alarms by enforcing equivalent importance on events regardless of their distance from the camera. We proposed a novel methodology using a depth-weighted loss function to diminish the overshadowing effect of closer objects, alongside a technique utilizing unusual training outliers to set the anomaly threshold. This approach was evaluated using video data from nine dementia participants across three cameras in a specialized dementia unit. The study found that the depth-weighted approach successfully lowered false positive rates and improved overall detection compared to the existing methods. It achieved the highest area under the receiver operating characteristic curve performances of 0.852, 0.81, and 0.768 across the three cameras. These findings are meaningful because mitigating distance-related sensitivity makes automated behaviours of risk detection viable for real-world deployment. Ultimately, this facilitates timely staff interventions, thus improving the safety and care of people living with dementia in understaffed facilities.