From Window Break Alerts to Person Detection Systems: Deploying Reliable, Low Power, Edge AI for Home Security Applications

This article is based on a webinar from Dr. Dave Garrett, Chief Architect and head of engineer at Syntiant Corporation. He’s a technical expert in building machine learning (ML) architectures in silicon and deploying ML solutions. In the webinar, hosted by SemiWiki, Dr. Garrett demonstrated how to successfully deploy edge AI neural networks using the Syntiant® ultra-low-power Neural Decision Processors™ (NDP). 

These NDPs sense, analyze and autonomously act to allow mission-critical and time-sensitive decisions to be made faster, more reliably, with minimal power consumption while offering greater privacy at the edge of the network. We’ve summarized the core takeaways of the webinar below. 

Deep Learning Enables Significant Reduction in False Alarm Events

False alarms on home security systems are the customers’ single largest complaint. According to recent studies, false alarms are the number one reason why people don’t actively use their installed home security systems. In these hardwired systems, there is no discerning and nuanced decision-making - they are either triggered or not.  

Syntiant’s solutions for home security applications are based on deep learning algorithms and aim to emulate the natural decision-making response of a person who can hear and sense an outcome, and make a determination about the underlying cause and event. Today, Syntiant offers three solutions for security applications, with the aim to expand in the future:  person detection (vision), window-break acoustic event detection, and door/window motion detection. For each application, the Neural Decision Processor™ is paired with a standard, low cost, sensor, and trained to detect the occurrence of a specific event, making these solutions turnkey and ready for deployment. 

1. Person Detection: NDP200 and Low-cost Camera

   
   Traditional home security systems use passive infrared sensors to detect the presence of a moving object. these sensors have no decision-making ability and rely on a single threshold, making them prone to false alarms: a god, a car, or even wind rustling the trees can trigger a ''person detect'' event.
   The Syntiant approach is completely different.: using a basic, low-resolution camera, Syntiant's NDP200 processor is trained to detect the presence of a person only and disregard all other objects. To allow the consumer security solutions to be installed in the most convenient and meaningful locations, many of these devices are compact and battery-operated. NDP200 is offer is a 5mmx5mm package and consumes less than 1mW of when used as an always-on, person detect processor. 

2. Window-Break Acoustic Event Detection: NDP120 and Microphone

   
   The second solution that Syntiant provides for security applications is to detect the sound of a window breaking. Just as with person detection, detecting the sound of a glass breaking is prone to false detects or missed detections. Syntiant’s NDP120 processor, trained with Syntiant’s own generated data models is exceptionally adept at distinguishing the sound of a window breaking from other household acoustic events.  The same processor, offered in a compact 3.1mmx2.5mm package, can also be programmed to trigger other specific acoustic events such as a smoke or fire alarms, alerting homeowners and security companies of a potential fire when no one is home. 

3. Motion Detection: NDP120 and Gyroscope

   
   The third Syntiant solution leverages the output of a six or nine-axis motion detector to determine events at the home’s entry points. For example, the system can be trained to detect the jiggling of a doorknob.

Syntiant’s Turnkey Solutions Encompass the Three Pillars of Deep Learning

Deep learning is ideally suited to be the new interface, enabling the natural interaction between the physical and the digital worlds. The challenge with deep learning is that to provide a total solution, three key elements are required: a processor where a neural network models are run; a large set of data assets, collected and classified, which is required to create the training datasets; and a training pipeline where machine learning data models are created.  Without any of these elements, the total solution is not possible. 

Facilitating the adoption of deep learning as the new physical-digital interface, Syntiant has been working on solutions that cover all three of these elements.  

1. Processor: Neural Decision Processor™

   
   Many deep learning solutions rely on traditional processor solutions that are not optimized for artificial intelligence or AI efficient. These solutions are large, consume a lot of power or are limited in their computational capabilities, making them not suitable for edge applications, where space, cost, and power consumption on a device is critically important.
   Syntiant has developed its own AI efficient™ (NDP) family of solutions.
   With native neural architecture, the NDP family offers 100x efficiency improvement and 30x increase in throughput, at 1% of the power and half the size, compared to commonly used microprocessor units (MCUs).
   Due to the unparalleled computational power for their size, processors in the NDP family do not need to send any computational data to the cloud, eliminating bandwidth challenges and security and privacy concerns. 

2. Data Platform

   
   One of the biggest challenges in deploying artificial intelligence and deep learning solutions is the scarcity of training data. In order to circumvent this problem, Syntiant has invested in developing its own platform. The Syntiant Data Platform automates the ingestion, labelling, aligning, cleaning, and synthetic data generation to turn raw data into training data sets. In addition, the platform generates synthetic data that can be used for customization purposes in easy use cases (e.g., wake words), or increase the robustness for complicated use cases.  

3. Training Pipeline

   
   Once a large enough data set has been collected and categorized, it can be used to create training data models for the neural networks. Generating production-worthy training models is an expensive and complicated process that includes augmenting the data to include many variations, mimicking real-world scenarios. Syntiant has designed and developed its own training pipeline that has helped to increase the robustness of application-specific solutions that they offer. Furthermore, the pipeline is not limited to use with the NPD family only and may be used to create production data models for other processors.

Smart Home Security Highlights Deep Learning’s Advantages for Edge Applications

Taking a closer look at how deep learning is applied to improve and expand applications in a smart home security system, Syntiant not only demonstrates an exciting use case of its specially designed Neural Decision Processors, but also outlines its vision for more intuitive, natural interface at the edge of the physical and digital world. 

Watch the full webinar on-demand here. 

About the sponsor: Syntiant

Syntiant Corp. is a leader in delivering end-to-end deep learning solutions for always-on applications by combining purpose-built silicon with an edge-optimized data platform and training pipeline. The company’s advanced chip solutions merge deep learning with semiconductor design to produce ultra-low-power, high-performance, deep neural network processors for edge AI applications across a wide range of consumer and industrial use cases, from earbuds to automobiles. Syntiant’s Neural Decision Processors™ typically offer more than 100x efficiency improvement, while providing a greater than 10x increase in throughput over current low-power MCU-based solutions, and subsequently, enabling larger networks at significantly lower power.