Label-Aware Temporal Modeling of Human Pose for Real-Time Fall Detection in Surveillance Video on Low-Power Edge Devices

Falls are a leading cause of injury among older adults and a common safety incident in workplaces and public spaces, yet the surveillance cameras that could witness them are rarely watched. Turning these cameras into fall detectors is harder than benchmark results suggest: real CCTV footage is low in resolution, often shows several people at once, and contains everyday movements that look like falls. We present a pose-based fall detection framework built for this setting and for inexpensive edge hardware. Each frame is processed by a lightweight multi-person pose estimator, people are followed by a tracker that keeps their identity while the body changes shape during a fall, and a causal dilated temporal convolutional network scores every tracked person from body-centric features alone, without appearance cues. Much of the improvement came not from the network but from the labels. Human-marked fall intervals turned out to be far shorter than the actual motion, with onsets misplaced by up to two seconds. We address this with label-aware supervision that extends positive targets, excludes ambiguous frames from the loss, and anchors each label to a person's box trajectory rather than to a tracker identity, supported by a keyboard-driven tool for precise person-level annotation. On a held-out test set of 158 unseen clips, the system reaches 98.1% clip-level accuracy, 96.0% fall recall and 98.0% alert precision, with a median alert latency of 0.74 s; five-fold cross-validation over 446 falls gives 96.6% recall. After INT8 quantization, the pipeline processes four concurrent camera streams of about one megapixel in real time on a single low-power edge board, without dropping frames.