NeNext: Edge-Based Video Intelligence for Human Safety Monitoring

Existing surveillance cameras offer a potential source of information about safety incidents, but continuous monitoring is difficult to maintain through human attention alone. NeNext is a project investigating practical, on-site video intelligence for identifying safety-relevant human events under the computational and operating constraints of edge devices. This paper presents a deliberately limited public account of its research objectives, evaluation design, and aggregate engineering observations. The reported internal study evaluated fall-event detection on 804 events from five cameras using a leave-one-camera-out protocol. A revised configuration achieved 86.2% event recall (693/804), compared with 85.3% (686/804) for a reference configuration. In a separate reduced-frame-rate evaluation, recall was 70.4% (566/804) versus 57.3% (461/804). Independent engineering records reported a reduction in one post-processing workload from 146 ms to 55 ms, and a two-stream observation recorded approximately 12 processed frames per second per stream over 180 seconds. These figures derive from distinct tests and should not be interpreted as an end-to-end comparative benchmark. The findings illustrate both opportunities and limits of edge-based safety monitoring, especially the sensitivity of outcomes to camera conditions and the need for longer false-alarm evaluations. Proprietary model design, software organization, and implementation methods are intentionally excluded.