Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{198673,
        author = {Aakash S and Dr. Ms. Asis Marceline V},
        title = {Accelerating Real-Time Fall Detection on Edge Cameras using TensorRT and INT8 Quantization},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {11},
        pages = {9984-9989},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=198673},
        abstract = {Building automated fall detection systems that ac-tually work in real-world hospitals is surprisingly difficult. The main culprit is the hardware bottleneck: modern deep learning models are incredibly resource-hungry. Typically, developers solve this by offloading the processing to the cloud. However, sending private video feeds over the internet creates a massive privacy liability, and any network lag during an emergency means the system fails when it is needed most. Trying to run heavy 32-bit floating-point (FP32) neural networks directly on edge cameras usually results in dropped frames and completely unviable inference speeds. To solve this problem, we engineered a decoupled edge inference pipeline that cuts the cloud out entirely. We started with a ResNet-18 model, exported it to a static Open Neural Network Exchange (ONNX) graph, and then compiled it using NVIDIA TensorRT. To bypass the strict VRAM limits on edge devices, we applied Post-Training Quantization (PTQ) using an INT8 Entropy Calibrator. By leveraging Kullback-Leibler (KL) Divergence, we mapped the 32-bit weights down to 8-bit integers without destroying the actual neural activation distributions. During testing, we also had to write a custom, dynamically balanced data loader for the UR Fall Detection Dataset to fix a major sequence leakage flaw we discovered. Ultimately, the pipeline shrunk the model’s physical storage by 74.10% (dropping from 42.72 MB to 11.06 MB) and achieved an inference latency of just 0.87 milliseconds per frame (1,150 FPS). This proves we can successfully run deterministic, real-time alerting directly on the edge.},
        keywords = {Edge Computing, Fall Detection, TensorRT, INT8 Quantization, Convolutional Neural Networks, Post-Training Quantization, MLOps, Triton Inference Server.},
        month = {April},
        }