Published Paper Details:

Rapid advances in client-side machine learning frameworks have made it feasible to execute deep neural network inference entirely within a web browser, without reliance on server infrastructure or specialized GPU hardware. This paper presents a unified, browser-based real-time object detection system built using React, Vite, and TensorFlow.js with the pre-trained COCO-SSD model. The system supports three distinct input modalities: live webcam streams, uploaded images, and uploaded video files. A key contribution is the accurate canvas overlay mechanism that dynamically computes scaling factors to ensure bounding boxes remain precisely aligned with detected objects regardless of display resolution or media size. The application further provides a real-time per-class object count panel aggregating inference results continuously across all modes. Operating entirely client-side, the system preserves user privacy, eliminates backend infrastructure costs, and enables zero-installation deployment. Experimental evaluation demonstrates inference speeds of 18-36 FPS for webcam mode on standard consumer hardware, with an overall F1-score of 92.4%, a confusion matrix validated per-class precision averaging 93.2%, and high bounding box alignment accuracy across all three input types.