UNLOCKING REAL-TIME INSIGHTS: TRANSFORMER-BASED DEEP LEARNING FOR VIDEO SURVEILLANCE AND HEALTHCARE

Abstract

Our framework is a deep learning model using a transformer to process video in real-time intended in surveillance and healthcare settings. We conduct a benchmarking of the performance of Vision Transformers (ViT) and Swin Transformers on three large-scale datasets, namely, UCF-Crime, VIRAT and MIMIC-CXR. On 2019-04-10 under the same settings, our models used 16x16 image patch and hierarchical attention mechanisms need to achieve the same mean Average Precision (mAP) as using CNN-based methods yet, our models increased by 12.4 percent. We fine-tuned the performance in the model using AdamW optimizer and carried out privacy-preserving preprocessing, like data anonymization. The Swin Transformer has been shown to achieve the highest trade-off between accuracy and latency, recording sub 100ms inference time, which is an acceptable limit of edge deployment. We examined too trade-offs among model complexity as well as responsiveness with a focus on feasibility of deployment. Ethical issues were taken care of with the differences of privacy approaches and federated learning models in order to protect sensitive information. We have established that Transformer models provide accuracy as well as efficacy to real-time video intelligence thus it is gratifying to conclude that it can be deployed in secure, mass scalability within the context of both public safety and healthcare.