The lack of quality onboard hardware analytics in surveillance cameras entails certain limitations and risks. Let's explore the development of cameras with integrated deep learning-based analytics and discuss the future of intelligent video analytics.

The field of deep learning originated from the concept of "artificial neural networks" in the 1980s. In the early years of this branch of artificial intelligence (AI), neural networks were modeled after the human brain, which is known to consist of more than 100 billion neurons.

A key limitation of early systems was the complexity of training the network. Hardware technologies were too slow to properly train a neural network that could solve meaningful real-world tasks.

Since 2000, the neural network research community has begun to attract the attention of industry labs to the work on deep learning networks. In recent years, real applications of this technology have spanned many fields, including handwriting recognition, language translation, automatic games (chess, Go), object classification, facial recognition, medical image analysis, the creation of fully autonomous vehicles, and many others.

According to data from market research leader Omdia, in 2020, 117 million professional video cameras were sold worldwide. 42% of them had a resolution of more than 4 megapixels. By 2025, the market share of such cameras is projected to reach 74%. At the same time, Omdia predicts that by the end of this year, the number of installed video cameras will exceed 1 billion. These cameras will generate a huge volume of information. The question arises: what to do with this impressive array of data? This information will only be useful if it can be processed and analyzed in a timely manner.

This is precisely the problem that deep learning algorithms help to solve, significantly enhancing the capabilities of hardware analytics on video cameras. With their help, devices can be taught to better filter out unnecessary data, which, in the world of big data, can save time, money, and human resources.

Currently, camera analytics are mostly limited to object detection and movement. This necessitates human intervention to identify objects and their actions using software on a central server. With increasingly sophisticated deep learning algorithms, cameras will be able to autonomously identify object types, the nature of their actions, and the appropriate response based on the data.

By 2025, it is expected that a new generation of chips will make built-in analytics on cameras more affordable. This could extend into the budget segment, which currently lacks deep learning cameras. As a result, the implementation of video analytics will become less costly in terms of finances and resources. The costs of servers previously needed and network bandwidth requirements will be significantly reduced. Deep learning-based intelligent video analytics on cameras will become as widespread as megapixel cameras are today.