Adding Cloud-Based Deep-Learning Object Detection Capability to Home Surveillanc...

Practical Deep Learning from Jupyter to Serverless Web Application

I recently installed a surveillance system equipped with four cameras and a Network Video Recorder (NVR) around my house. Unfortunately, almost all false alarms were triggered by moving plants or tree shadows or squirrels. None of these alarms can be filtered out by traditional image processing capabilities coming with the system.

Like most deep learning practitioners, I know object detection programs can filter out these false alarms. But they either require an expensive commercial contract or a computer on my home network. Since I want to keep the cost low, having a computer seems the right choice. However, it’s still a rather large initial capital investment plus the recurring 24/7 electricity cost. The computer also requires setup, maintenance, and shelf space. Its fan noise or heat dissipation from the closet is another nonsense I prefer not to deal with at home.

Most false alarms are simply trigged by moving tree shade and plants. These false alarms cannot be filtered out using traditional image processing techniques such as adjusting contrast threshold or setting active zones

Upon further research, I found out using serverless web APIs is the best solution. It not only gives fast response but also charges a very small fee based on usages. I also want to optimize the deep learning algorithm by myself or to reconfigure the implementation for advanced deep learning applications. I have thus chosen MXNet running on AWS. The combination allows easy deep learning code development using Jupyter, optimized library performance, abundant pre-trained models, and the powerful open cloud infrastructure.