Thinking about where we are now versus where we are going is something that can help us. These days, there is still a lot of discussion about what is possible in the future, compared to the products that are physically available on the market. Much of the talk focuses on the following topics:
- Practical ways to use deep learning and neural networks
- How these techniques can improve analytics and significantly reduce false positives for events that are of high importance.
Talking to users about analytics, it becomes clear that many of them are still not fully utilizing these analytics. In some cases, this may be due to past experiences with previous generations of analytics, which were full of false positives, but in other cases, it may be something else; It may not occur to people that there are reliable analytics that can meet their specific and unique needs. The good news is that with the help of AI technology, or more precisely, deep learning technology and neural networks, we are reaching a new level of advanced analytics and data collection in two key areas:
1. More precision in existing analytics: In the past, system developers have struggled to define:
- What is a subject?
- What is motion?
- What is interesting motion that we are interested in tracking, and what is random motion that we should ignore?
A good example is the wind blowing through trees, or blowing a plastic bag in the air. Something as simple as a long-term motion detection mechanism is often confused by false positives caused by wind. Users can try to lower the system’s sensitivity for light breezes, but as soon as a major storm approaches, the system’s motion events are triggered. Using neural networks and deep learning to define subjects such as people, cars, buses and animals means that traditional analytics can now focus on subjects. These new algorithms are trained to recognize individuals by seeing images of thousands of people. This repeated learning is exactly the same way that a neural network can learn to recognize a person or a car. By learning these examples, the system can apply this advanced knowledge and information to existing analytics. For example, if a car crosses a line at an entrance gate, it can acknowledge the movement and not block it. However, if a person crosses the exact same line, we may want to send an alert. Shadows should be ignored and the movement of trees should not be taken into account. A car driving in the opposite direction certainly requires an alert, but people moving freely are fine. All of the old motion-related analytics, including appearing/disappearing, crossing the line, abandoned objects, and loitering, will become more accurate and capable of further refinement using AI and subject recognition.
2. Better data mining: With AI, cameras can tell us more about the subjects they detect. It’s not just that the system can recognize a person, but that the person is wearing a green shirt and black pants, has dark glasses, and is driving a small red sedan. This extra information is embedded in the metadata recorded with the video (metadata is data about data). Each frame of video has its own header, which contains extra analytics metadata. This means that this metadata is time-aligned with the video. If this data is stored where the video was recorded, you can look at the recorded files and look for a green shirt in this metadata. This can reduce hours of searching to minutes or less.
Edge vs. Server
Analytics can be run on a dedicated server or on the edge in the camera itself. Server-side AI (server-side refers to the set of operations that are performed on the server side of a client-server network) is used when more data-intensive analysis is required; such as large-scale comparisons across databases, which are common in facial recognition, ALPR, etc. However, even for computationally intensive tasks, the efficiency of processing speed and low bandwidth requirements can be met by using a hybrid approach, which involves edge-based devices and servers working together. Instead of sending raw video, the metadata obtained from the AI at the edge can be sent to a server-side application; Raw video that forces the server to decode multiple video streams to perform just one set of analyses.
With these new AI capabilities on the edge, security cameras are more powerful than ever, and the accuracy of analytics has increased exponentially.
Source: Security Magazine