<![CDATA[Real-time monitoring system development frequently encounters accessibility challenges when deployed on mid-to-low-end hardware. This study documents the experimental process of developing a web-based AI object recognition system utilizing TensorFlow.js and the COCO-SSD model. Prior research employing COCO-SSD has demonstrated suboptimal performance, with response times exceeding 33 ms. During the development phase, custom logic incorporating overlap mechanisms and cooldown features was implemented to address limitations inherent in basic object detection when recognizing human-object interactions. To optimize real-time performance, this logic was applied at the application level rather than within the AI model itself, leveraging the latest deep learning methodologies proven to outperform YOLO. Using a private training dataset comprising limited facial and indoor object images, the system successfully visualizes bounding boxes and sends instant WhatsApp alerts via whatsapp-web.js. The methodology adheres to an integrated web-based object detection workflow. Experimental results demonstrate a responsive system with latency below 30 ms, meeting real-time performance standards. This paper concludes that the JavaScript-based AI stack, combined with spatial logic, effectively provides a functional solution for automatic activity recognition.]]>
Copyrights © 2026
