How deep learning and AI can deliver value to industries

Words such as big data and the IoT have been hot topics in the last few years. Better image resolution and the IoT have created a massive amount of data, which makes sifting through and maximizing it much more daunting. To meet these demands, camera and software manufacturers are incorporating AI and deep learning technologies into their analytics. These advances in technology are helping to tackle all this information, thereby increasing the overall effectiveness and efficiency of such systems.

Using Deep Learning and AI in Security

In security alone the AI market was valued at roughly US$3 billion in 2016, according to MarketsandMarkets; it is expected to reach $34.8 billion by 2025, at a CAGR of 31.4 percent during the period 2017 and 2025.

“Today’s security industry has reached a critical mass in the volume of collected data and the limits of human attention to effectively search through that data,” said Willem Ryan, VP of Global Marketing and Communications at Avigilon. “As such, the demand for video analytics is increasing globally. Through the power of AI, Avigilon is developing technologies and products that dramatically increase the effectiveness of security systems by focusing human attention on what matters most.”

Avigilon’s Appearance Search technology, a sophisticated deep learning AI search engine, is designed to make searching surveillance footage easy. “It incorporates the unique characteristics of a person’s face, as well as shape, size, texture, color, clothing and accessories, to quickly locate a specific person of interest across an entire site,” Ryan explained. “Using the unique combined features of a face based on hair color, texture and other facial features and accessories, this technology is designed to increase the speed and accuracy of investigations by detecting and understanding that it is searching for the same person, even if items such as their clothing change over time.”

AI is also being used to bring a new level of automation to surveillance. An example is Avigilon’s Unusual Motion Detection (UMD) technology. “Without any predefined rules or setup, UMD technology is able to continuously learn what typical activity in a scene looks like, and then detect and flag unusual motion. This allows operators to search through large amounts of footage faster, as UMD focuses their attention on the atypical events that may need further investigation, helping to reduce hours of work to minutes,” Ryan added.

Tom Edlund, CTO of BriefCam explained how the use of deep learning helps them achieve significantly higher accuracy than traditional computer vision algorithms, provided there is sufficient data for training and enough resources to compute. “BriefCam offers applications for rapid review and search, real-time smart alerts and quantitative analysis of video from surveillance cameras. Our video processing module uses deep neural networks for object detection, classification and recognition.”

“We use deep learning to learn the video scene dynamics and to generate rich and unique metadata. To give an example, a user can search for a woman, 1.7-meters tall wearing a red dress across multiple cameras. The video processing works on all stationary cameras without need for manual configuration,” Edlund added.

Creating More Efficient Buildings With AI

AI is also being used to create more efficient buildings. Things like energy efficiency are becoming increasingly important as smart cities and smart buildings adoption continue to grow.

However, smart buildings can be made even smarter with the help of AI and deep learning. An article by an analyst at Navigant Research named three reasons for why commercial buildings are ready for AI: rapid adoption of IoT opens the door to AI in buildings, customers expect technologyenabled convenience, and AI can bolster building owners’ top line.

Verdigris Technologies offers an AI-based IoT energy sensor and software as a service (SaaS) for commercial and industrial buildings that learns equipment electricity consumption and optimizes existing building management systems.

“Our hardware solution generates a highly granular data stream (monitored at 8kHz). This data provides the foundation for AI models,” said Jonathan Chu, CTO of Verdigris Technologies. “We use deep learning techniques to provide customers with insights into their energy usage, otherwise challenging to achieve through regular utility smart meter data.”

Previously, the company used a physicsbased machine learning technique for device detection, but now uses deep learning. “This change enabled better results for forecasting and disaggregating energy loads. The forecasting algorithm is a supervised recurrent neural network model using long short term memory on historical data.

Disaggregation is a hybrid model combining an unsupervised recurrent neural network and a fully connected network with the capability to turn on layers. Classification is on our product roadmap is to use deep learning as well,” explained Chu. “By wirelessly collecting high fidelity data, we are able to apply deep learning to provide forecasting and disaggregation solutions on an accessible, real-time basis.”