Behavioural and psychological symptoms of dementia, including agitation and aggression, pose substantial health and safety risks in residential care environments. The widespread use of video surveillance in common areas within these facilities offers an opportunity to develop automated systems for detecting behaviours of risk. Such systems can provide real-time alerts to staff, facilitating timely intervention and helping to prevent the escalation of potentially harmful situations. A major challenge for these systems is their adaptability to new environments without a significant drop in their detection performance. We propose noiseCAE to tackle this challenge by utilizing partial masking of scene background to mitigate scene bias and assist the autoencoder model to learn scene-invariant normal behaviour characteristics. This approach helps the autoencoder to generalize better to unseen camera scenes without any additional training. The data from nine individuals with dementia, recorded using three cameras positioned in different hallways of a dementia care unit, was utilized for this study. The generalization performance of noiseCAE was investigated in a cross-camera setting, where noiseCAE performed better than the existing method for three out of six cases. This motivates further research to develop scene-invariant cross-camera behaviours of risk detection systems for people with dementia in care environments.