Currently, AI has become a hot keyword in various industries, and the physical security industry is no exception. Artificial intelligence is also an important force that has the potential to change and reshape the industry.
In a broad sense, artificial intelligence (AI) refers to the intelligence of machine computation, not human beings themselves. In the security industry, AI refers to the technological application of machines that mimic human and other biological cognitive functions, that is, imitating the human brain's thinking and methods for learning and solving problems.
Artificial intelligence drives the rapid development of the security industry
Currently, three major trends in the computer industry are driving the rapid growth of artificial intelligence. These three trends are:
First, the rise of computer hardware has enabled the handling of complex computations, especially graphics processing units (GPUs, which use a "parallel processing" model instead of the familiar "serial processing" model of CPUs). Multiple computational tasks can be processed simultaneously in parallel, far more efficiently than the "serial" model. Moreover, this is a more scalable approach: breaking down large problems into many smaller problems that can be solved concurrently. Second, the development of more efficient methods for "training" systems, particularly neural networks, has enabled them to work in parallel with GPUs.
A neural network is a computational system composed of many simple, highly interconnected processing units, typically structured in layers, each consisting of interconnected nodes. The results computed by each layer determine the input to the next layer. Neural networks can have more than a hundred layers, thus enabling them to process vast amounts of complex data. Third, the proliferation of sensors (including cameras) generates enough data to allow systems to be effectively "trained" (e.g., "big data").
The surge in "big data" has provided a massive amount of training data. Security video surveillance data accounts for 60% of big data and is growing at a rate of 20% annually. This surge in data provides a driving force for the development of artificial intelligence and enhances the functionality of systems.
Artificial intelligence system training
In neural networks running on GPUs, the learning rules continuously optimize and adjust the weights (importance) of the connections; each layer has different "weights" that reflect what was learned in the previous layer. When presented with a data model (such as a video image), the neural network can determine what it might be by analyzing patterns.
Training involves determining initial and final results and adjusting connection weights appropriately. In highly general terms, this is how AI systems "learn." However, the entire "training" process is divided into multiple stages, like filters, with the results of each stage guiding the path to correct analysis.
Deep learning is a type of broader machine learning approach and is the concept most relevant to the security video industry. Deep learning requires the use of large amounts of data (e.g., video images) from neural network learning systems.
Deep learning in video surveillance systems
Interconnected processing units in neural networks work in parallel with GPUs, designed to mimic the human brain's analysis and processing of problems through billions of neurons. Artificial intelligence, particularly deep learning, is becoming the foundation of next-generation video surveillance systems, giving traditional systems superior performance.
This approach dramatically changes the effectiveness of video surveillance systems. Previously, computers were programmed with video analysis algorithms. In contrast, deep learning systems are far more "trained." If you want to identify a cat, you provide a large number of images of cats, and the system breaks them down into smaller components and looks for common data. It then "learns" the shared features in these cases.
To maximize training accuracy, the more data the system is presented with, the more precise it becomes, meaning it "learns" more. Through learning from a large amount of case data, deep learning systems develop corresponding recognition patterns.
From Training to Inference
While computer programmers can spend months writing instructions to tell a computer what a car looks like, neural networks can "learn" by showing a large number of examples without requiring programming. Furthermore, training a neural network is also time-consuming, potentially taking hours or days. Training is a computationally intensive task.
However, once the neural network is trained, we can use it to "infer" and assist in decision-making, such as determining whether a cat is in a newly captured image. This allows us to deploy trained systems on devices such as network video recorders (NVRs), and even in video cameras at the network edge, enabling them to quickly identify target objects and make corresponding decisions rapidly.
Deep learning can achieve superhuman pattern recognition accuracy and is also resistant to interference, capable of classifying and recognizing thousands of different features. For example, the latest facial recognition and license plate recognition systems have accuracy rates approaching 100%. These characteristics make deep learning highly valuable and significant for video analytics applications.
Disclaimer: This article is a reprint. If there are any copyright issues, please contact us promptly for deletion (QQ: 2737591964). We apologize for any inconvenience.
