Applications of deep learning technology for video surveillance

The artificial intelligence field often referred to as cognitive computing or machine learning has recently become extremely well-known. The rapid growth in the field of "deep learning" technology in the last several years has proved significant in a variety of sectors. The major companies in the field of industry like Google, Microsoft, Facebook, IBM and many others have made huge amounts of capital into the field of artificial Intelligence. Machine learning has been exploding on the scene due to the advancement in emerging "deep understanding" technology.

Deep learning developments can have implications for the security industry in general also. For video dataset analytics, for instance deep learning has shown promise of improving some difficult problems, though more work is required. This article will discuss the advancing field of deep learning as well as its possible impact on security and surveillance market.

The evolution of deep learning

The area of deep learning developed out of "artificial neural networks" in the 1980s. In the beginning, this particular branch of artificial intelligence, neural nets are modeled on the human brain, which comprises more than 100 million neurons. This field never developed in the 1980s and the '90s for a variety of reasons. The major drawbacks of early systems were the inability to train the neural network and the CPU technology on hardware were not fast enough to develop a neural network capable of solving real-world problems.

The 1990s and 1980s were the darkest years of neuronal research. Since 2000 the research community of neural networks has begun to attract industry labs' attention due to the advances in deep learning research in academia from such institutions as the University of Toronto, NYU, Stanford and others. In the last few years, the real-world applications of deep learning cover a wide range of fields, such as handwriting recognition, translation of languages and automatic gaming (chess/Go) playing objects, face recognition, medical image analysis autonomous driving vehicles and many other areas.

One of the examples of the excitement associated with deep learning is the latest breakthrough by Google's AlphaGo program, a computer program that first time defeated an expert human Go player on October in 2015. The level of sophistication of AlphaGo's capabilities Deep learning program has shocked many of those working in artificial intelligence because of how complex the old Asia GO game, which is a lot more complex than Chess.

Applications for video surveillance to aid in deep learning

While the use of deep learning is utilized in various industries and has produced breakthrough results in comparison to traditional systems but not all applications can benefit from deep learning. In the realm that of surveillance through video many applications stand out to profit from the use of deep learning.

Face recognition: The use of deep learning has greatly increased the accuracy of facial recognition. In the National Institute of Standards and Technology (NIST) has conducted the Face Recognition Vendor Test (FRVT) test for the last decade. The advancements over the past 20 years in the accuracy of face recognition have been reduced to three times the rate as per the NIST Interagency Report. The majority of the top-performing commercial face recognition applications are made up of deep learning. The accuracy is as high as 99.9 percent for controlled environments such as facial recognition for immigration at airports as per research conducted done by Facebook as well as Tel Aviv University.

The detection of objects and persons: The detection of objects and people is another field where deep learning has made huge growth. In the past five years the IMAGENET database has organized "large scale visual recognition challenge, "large large-scale visual recognition contest" wherein algorithms that use image data collection are asked to recognize, classify, and identify a database comprising over 150,000 photos sourced via Flickr as well as other engines. The data is categorized in 1,000 categories of objects. A variety of deep learning algorithms are trained by using more than 1.2 million of images in the IMAGENET databank that are running on GPU accelerators that use hardware. The accuracy gains vary from 72% to over 90% between 2010 until 2014. In 2015 the majority of IMAGENET contestants employed deep learning techniques.

Deep learning , in its early stages has shown lots of promise in tackling difficult video analytics issues

In the realm of surveillance video, a number of applications stand out which can benefit from deep learning.

Deep learning-based video surveillance systems

The main benefit that deep-learning algorithms have compared to traditional computer vision algorithms is that a deep learning systems can be continually improved and trained using larger and better datasets. Numerous studies have demonstrated that deep-learning systems can "learn" to attain 99.9 accuracy on specific tasks, as opposed to computer algorithms that are rigid, where it is extremely difficult to enhance an algorithm beyond 95% accuracy.

Another advantage of the deep learning is "abnormal" detectability of events. Systems that use deep learning have demonstrated incredible ability to identify unpredictable or undefined events. This technology has the possibility of drastically reducing false positive detection incidents that are a problem for the majority of security video analytics systems. The inability to reduce false positive detection rate is the main issue in the video surveillance industry which has prevented the widespread acceptance of a variety of vendors' intelligent video analytics systems.

Open issues and deep learning technology for video

In its early days, deep learning has shown plenty of promise for improving difficult and difficult video analytics issues. There is still a lot of work to be done in order to improve the general deep-learning system to recognize and understand particular events in the domain that are specific to security-focused environments.

The other issue is that the engineers in deep learning technology is in extremely shortage. Many of today's graduates come from top universities, and after graduation, they are taken over by Internet giants such as Google, Facebook, Amazon, Microsoft, etc. The competition for skilled computer scientists is fierce.

The third issue is the reality that not all algorithms for video analysis are suited to be used using deep learning. There are a variety of older computer vision algorithms created over time which are extremely well-suited and used in highly successful commercial products. For instance license plate recognition works extremely well when using computer-vision-based algorithms. Industry must conduct more research into hybrid systems that blend the very best computer vision algorithms with deep learning.

How GTS can help you?

Global Technology Solutions understands the need of having high-quality, precise datasets to train, test, and validate your models. As a result, we deliver 100% accurate and quality tested datasets. Image datasets, Speech datasets, Text datasets, ADAS data collection and Video datasets are among the datasets we offer. We offer services in over 200 languages.