AI Video Surveillance (Finally) Goes Mainstream In 2020By John Honovich, Published Sep 03, 2019, 10:36am EDT
While video surveillance analytics has been promoted, hyped and lamented for nearly 20 years, next year, 2020, will be the year that it finally goes mainstream, marking a major advance for the industry and security users.
In this report, we explain why this is important, what will happen, why it will happen and what impact it will have.
Why This Is Important
Alerting and searching is made much harder when systems cannot determine what is a person vs what is leaves, shadows, headlights, birds, etc. Alerting cannot be trusted since so many of the alerts will be obviously false and searching is more difficult since users are forced to wade through so many clearly irrelevant results.
This will solve a long-standing frustration of surveillance systems - that there is so much video but no way to manually 'look' at it all. Instead of doing so, AI will automate that process, making systems significantly more valuable and increasing demand for systems.
What Will Go Mainstream - People, Then Vehicles
"AI" is an immense field. To be specific, what we are confident will go mainstream in 2020 is the ability to reliably detect what is a person vs what is not. Vehicle detection will follow.
Harder Not Mainstream
Harder analytics, like facial recognition, tracking people across cameras, detecting smaller objects (like guns, knives, backpacks, etc.) will not be mainstream in 2020. There will certainly still be many areas where companies can differentiate on harder "AI".
Mainstream Means Widely Offered
Obviously, some companies have offered "AI" / video analytics that have worked for years, however, this has remained a distinct niche. For example, see the (mostly) terrible results from our 2018 professional video analytic shootout.
The difference in 2020 is that it is going mainstream, meaning that almost all manufacturers will be offering fairly accurate and reliable people detection analytics. And the ones that do not are going to stand out as outliers, the other way.
20 Years In the Making, Nearly
What is wild about this is that for nearly 20 years, companies have been claiming to deliver working video analytics. For example, it has been more than 16 years since Object Video's analytics won 'Best In Show' (one of a series of 'best' failures).
Video analytics has been hyped for so long that that 15 years ago, most industry professionals viewed analytics as being more promising than IP cameras. Of course, the last decade has shown otherwise, with IP / MP surging and analytics dealing with a decade of despair.
Why 2020 Is Different
This time really is different, due to (1) deep learning and (2) it being productized in hardware and software kits that make it easy for this to go mainstream.
The benefit of deep learning is that, in the past, video analytics relied on heuristics (like what is the aspect ratio of the pixels changing and how long has it lasted, etc.). These heuristics were often, literally, shots in the dark and were prone to major errors. With deep learning, systems are trained with examples of what is a person (e.g.) and what is not.
The downside of deep learning is that historically it has been very resource-intensive (think servers filled with GPUs).
What is making deep learning go mainstream for AI video analytics is that it is being productized to fit with video surveillance cameras and recorders. On the hardware side, there is Intel Myriad (Movidius), Ambarella CV chips, Huawei Hisilicon, Qualcom and others providing the chips that make it much turn key to deliver AI on these devices. On the software side, the most notable is XNOR.ai, the inventors of YOLO, who have built a business providing optimized software modules to work even on commodity hardware
How We 'Know'
We 'know' and can be so confident because we run the world's leading video surveillance testing program and are already seeing it. The most stunning example is the $20 Wyze cam (using XNOR.ai) that is delivering viable person detection analytics. Another example is Dahua Analytics+ (tested). While Dahua has been terrible at video analytics for years, new hardware has enabled them to deliver viable person detection analytics.
Beyond that, we have information on development plans of various companies and top of the roadmap for most manufacturers is the very obvious inclusion of chips and/or software modules to make deep learning work on their devices.
This will be a significant boost to the industry. The previous era, the shift to megapixel, has largely run its course and for the last few years, along with the now-ended race to the bottom, the industry has lacked a clear technological driver.
Now, end-users will legitimately find significant value in AI video analytics that can help them improve alerting and searching. There will be genuine value in buying new cameras, recorders, servers or software to get alerts on potential intruders and speed up the process of conducting searches.
Challenges Remain In Selection
While this will go mainstream, there will still be challenges in the selection, such as:
- Ensuring that the analytics will be accurate in own's environment. While false alerts will be phenomenally reduced compared to the previous era, evaluating and determining which works best in one's environment will still be important.
- Evaluating how well these analytics work within one's existing equipment and monitoring software (VMS, NVR, central station, PSIM, etc.) will be important as analytics will vary in how easy to use and deeply integrated they are.
- Determining what type of advanced analytics one wants will remain an important consideration. While people detection will increasingly be as expected as H.265 or IR, etc., AI can deliver more advanced classification and decision making, which will remain much more of a differentiator.
Questions / Thoughts
What do you think about AI video analytics going mainstream? Agree? Disagree? Specific Questions? Let us know in the comments.
8 reports cite this report:
Back to Top