This work aims to evaluate a non-wearable fall-detection system for indoor activities of older adults. The computer vision proposal uses a Convolutional Neural Network (CNN) architecture using the lightweight SSD-MobileNet-V2 model. Transfer learning techniques were applied to develop this approach. During the training process, a combined dataset conformed by the UR Fall Detection and an own dataset was used. Verification tests were performed under diverse scenarios, participants, and platforms to provide greater reliability to the evaluation results. Different distances from camera to person, lighting levels, attire, participant’s gender, and age were considered. For evaluation purposes, two platforms were selected: one based on a Single Board Computer (SBC) and the second on a laptop computer. Results show an accuracy of \(95.5\%\) , a precision of \(99.4\%\) , a sensitivity of \(91.6\%\) , and a specificity of \(99.5\%\) when executing the approach on a SBC. On the other hand, when the approach runs on a laptop, it achieves an accuracy of \(96.6\%\) , a precision of \(97.3\%\) , a sensitivity of \(96.0\%\) , and a specificity of \(97.4\%\) . Lastly, the system sends an alert notification when a fall is detected through a messaging platform. This event supports effectively medical assistance in reducing fatal consequences, especially for older adults living alone.