When smart device startup Alcidae released its Panoramic AI camera in December 2018, Chinese e-commerce customers instantly began snapping it up. In addition to a 1080P HD camera with 360-degree panoramic view and 100-degree vertical rotation, the device comes with a host of AI home security features that similar products on the market could not match. In the space of a month, it shipped 100 thousand units and topped the Top-10 smart cameras list on JD.com, one of the two leading e-commerce platforms in the country.
Alcidae’s Panoramic AI camera, powered by HiLens
Underlying the AI security features of Alcidae’s smart camera is an IaaS + AI algorithm solution provided by HUAWEI CLOUD. Now called HiLens, Huawei and Alcidae originally began working on the solution in April 2018, with Huawei proceeding to unveil it at HUAWEI CONNECT that October.
HiLens allows Alcidae’s device to offer a broad range of AI and cloud-computing powered features, including facial recognition and infant cry recognition, in a market where most products do little more than allow the user to view the raw footage from their smartphone and other devices.
The success of Alcidae’s device on its release two months after HUAWEI CONNECT is a clear indicator of the market appetite for smarter functionality among home security devices. Still, London-based analysts IHS Markit say that even facial recognition and voice recognition are merely scratching the surface of the capabilities that AI has to offer the home security market in 2019.
And they’re right: the full extent of AI capabilities available today far exceeds those represented on the market, including even Alcidae’s feature-rich product.
Cutting-Edge VS Cutting Costs
The reason for the slow adoption of AI capabilities among smart home products, however, is not complacency on the part of enterprises. Rather, the issue lies in the high cost of commercializing cutting-edge AI features.
The maximum one could expect to charge for a home-use security camera in China without pricing out the average consumer is around US$73 (RMB 500). Alcidae’s Panoramic AI camera sits comfortably in the middle of the price range at around US$44 (RMB 299). This price range places major restrictions on the caliber of AI capabilities available to the device.
One such restriction is the cost of cloud computing. While powerful and accurate machine learning algorithms can easily be supported by the practically limitless cloud computing capabilities available today, these capabilities do not come cheap – especially when you consider how frequently a smart camera, for example, needs to run a facial recognition algorithm.
Meanwhile, the alternative – running AI algorithms on the device side – comes with significant compromises in terms of performance. Chips with relatively low computing power and a limited amount of memory are all that can be affordably placed in these devices, leading to simplified algorithms which produce high false-positive rates in recognition tasks.
Commercializing AI with Device-Cloud Synergy
To enable Alcidae to provide multiple, high-performing AI features at an affordable price for consumers, HUAWEI CLOUD developed their HiLens solution based around the concept of “device-cloud synergy”. The idea is to allow devices to flexibly make the most effective use of the computing capabilities available on the device and in the cloud to achieve the best performance-cost ratio.
HiLens realized device-cloud synergy through a combination of proprietary algorithms, Ascend chips, a cloud platform, terminal hardware, and other full-stack capabilities. In Alcidae’s case, it achieved a massive 10x performance boost for their AI features at a tiny fraction of what it would cost to run these same features entirely in the cloud.
Two key features that benefit from cloud-device synergy are facial recognition and video analytics. In addition, Huawei was able to implement some features fully on the device side.
Case 1: Facial recognition
Alcidae’s smart camera uses facial recognition in a number of features, including intruder detection and intelligent video retrieval, which allows users to quickly search camera footage for tagged faces, strangers’ faces, and other content.
Due to the limited computing power available on the device side, the chip-based facial recognition algorithm is cropped, limiting its recognition accuracy. By comparison, the corresponding cloud-side algorithm offers much greater complexity and accuracy, but is too expensive to use frequently.
Device-cloud synergy is implemented here by means of primary recognition on the device side supported by secondary recognition in the cloud. When the device-side algorithm recognizes a face, it calculates a confidence score and compares it against a threshold – typically, 93%. If the image does not meet the confidence threshold, it is sent to the cloud for further processing using the more powerful algorithm.
95% of images are capable of meeting the confidence threshold, meaning that the vast majority of facial recognition is performed entirely on the device side with no cloud computing costs incurred. Meanwhile, leveraging cloud computing capabilities for the 3–5% of borderline cases allows Alcidae to achieve an overall accuracy level equivalent to running the entire feature in the cloud, but at a fraction of the cost.
Device-cloud synergy allows Alcidae to achieve a facial recognition accuracy level equivalent to running the entire feature in the cloud, but at a fraction of the cost
Case 2: Video analytics
Unlike most smart home cameras which simply store footage in the cloud for the user to look through manually, Alcidae’s device supports a versatile search function with eight search criteria, helping users to quickly home in the footage they are looking for. This function is geared towards use cases such as parents who want to quickly access all footage of their child from throughout the day.
This search function is made possible through cloud video analysis. Cloud video analysis is most widely used in the public security sector, where video footage is first stored in the cloud and retrieved only when required, e.g. during an investigation. Retrieved footage is decoded before being analyzed frame by frame – a process that requires vast computing resources.
This model is unsuitable for the home security sector, where customers expect to be able to access much more of the stored footage but are unable to bear the high costs of cloud-based analysis.
In this case, device-cloud synergy was implemented by dividing the video capture into two streams, one being sent to the cloud for storage while the other stream undergoes analysis on the device side. The results of analysis are then sent to the cloud, which tags the footage accordingly.
For example, when a clear face is detected by the device algorithm, the image is sent to the cloud, where it is matched to the corresponding point in the video stream and the timestamp is recorded. The fact that only a limited set of faces selected by the user are sought for during analysis further reduces the costs.
Case 3: device-only infant cry recognition
Complementing the device-cloud synergy approach, some features – including human figure recognition and infant cry detection – are optimized to run entirely on the device side, without compromising on performance.
The biggest challenge in this regard was the limited memory available on the device. With a total capacity of 64 MB and just 5 MB left after system algorithms are accounted for, Huawei had to refactor the infant cry recognition algorithm for chip implementation as well as compress and optimize the machine learning model. Finally, the infant cry recognition feature was realized within 3 MB.
Comparable features in similar products are typically plagued by low recognition rates and high false positives due to the complex audio environment of the home. Meanwhile, Alcidae’s product offers highly effective cry recognition within a 4-meter radius and has a false positive rate of just 5% – a milestone for a notoriously challenging feature in the industry.
Infant cry recognition can also be used to trigger alarms or notifications for the user, allowing parents to respond to their child more quickly in potential times of need. It is also used for intelligent retrieval of recorded audio, similarly to the way facial recognition supports this function for video.
Empowering IoT Solutions in the Home and Beyond with HiLens and Ascend
As IHS Markit have pointed out, there is still a long way to go before the AI home security sector matures completely. In the near future, HiLens is ready to support features such as fall detection (for use in elderly care scenarios), pet identification, recognition of other household objects, and more. Features like this are in line with the current trend of the household security category broadening to cover other aspects of care in the home.
Longer term, the home security sector will serve as a proving ground for HiLens device-cloud synergy solutions, preparing them for implementation in wider scenarios.
Currently, campuses and shopping malls are two major target scenarios. In the campus scenario, the main applications for smart cameras are campus security and management, with functions including license plate recognition, vehicle identification, and face recognition.
Huawei’s recently developed AI chips will help support diverse application of HiLens in future by providing additional computing power on the device side. The Ascend 310, for example, is an efficient 12nm SoC (System on a Chip) designed for low-power computing scenarios. It has a power consumption of 8W and a performance of 8 teraFLOPS under FP16 and 16 teraFLOPS under IN8.
The Ascend 310 is currently in mass production, and will be commercially available by June 2019.
Find out More: Explore the HiLens IaaS Platform
Developers and partners can access the HiLens platform, which provides end-to-end capabilities for feature development, deployment, and management in the cloud.
HiLens provides a Skill Market for developers to quickly release skills they have developed for use in their industries. This year, an initial batch of 200 Huawei-developed skills will be released to the Skill Market for demonstration purposes and to stimulate market development with partners. On the HiLens platform, developers will have access to easy-to-use components for secondary development, enabling rapid customization of industry applications.
Like Alcidae in developing their Panoramic AI camera, developers looking to leverage HiLens to create powerful AI products at competitive prices can rest assured they will be able to benefit from the wider Huawei ecosystem: technical support from HUAWEI CLOUD; the computing power of Ascend chips; the Huawei “HiLink” Smart Home Ecosystem protocol; Huawei’s terminal companies; marketing support; and more