Instruction vs. Deduction: Deep Learning and Advances in VCA
Written by Brian Carle, Director of Product Strategy at Salient. Published in Security Today Magazine
As camera counts and the data they provide grow ever-larger, it becomes increasingly difficult for organizations to monitor, perform investigations, and draw useful conclusions from the valuable information gathered by their video surveillance infrastructure.
Video analytics have long been seen as a technology solution to help identify activity and information from all the video data. Video analytics have largely fallen short of delivering on that market expectation. However, Deep Learning may change that. But what is Deep Learning, and how can it improve on conventional techniques?
Machine Learning Techniques and VCA
Most Video Content Analytics (VCA) developed to-date have been based on traditional, algorithmic, Machine Learning techniques. Deep Learning is a more advanced evolution of machine learning, using sophisticated, artificial neural networks.
In the context of VCA, both Machine Learning and Deep Learning instruct software to develop a model of objects based on a variety of attributes the software “learns” about those objects. The model helps the software to later identify and categorize an object in the video feed which matches those attributes the software has learned. For instance, an object moving through the camera’s field of view may be taller than it is wide, as opposed to another object, which is wider than it is tall. The VCA software may classify the first object as a person and the second as a vehicle, based on those attributes.
In reality, multitudes of data points are used to classify objects, but some attributes are more important than others. The VCA software will weigh the various criteria it uses to classify objects in order to determine the probability that an object is more likely to be a person, vehicle or something else. Once an object enters the scene, the object is analyzed, and its properties are measured. To determine what an object is, the VCA may begin by looking at the object’s dimensions, color variation, and movement patterns.
For example, the software determines the object is wider than it is tall, is primarily red, and is moving at a relatively rapid pace in a single direction. Based on these observations, those attributes are compared to the existing model of what properties represent a car, a person, or other objects. Based on the comparison against existing models, the VCA software finds the object is 88 percent likely to be a vehicle, seven percent a person, and 22 percent “other.” The object is identified as a car, and data is collected along with the video which may allow the user to later perform a search on all red cars travelling from left to right in the scene.
Limitations of Machine Learning Analytics
Machine Learning creates a model of an object based on data fed to the program by its developers. This data is compiled by people, and will therefore be inherently limited to the set of attributes a developer chooses to collect and feed to the program.
To continue with the “person versus vehicle” example, an object may be classified as a person by Machine Learning VCA if the dimensions of the object show a greater height than width, as opposed to a vehicle, which may be wider than it is tall. Given those criteria, VCA classification may fail in the case of a person crawling through a scene, or a person carrying a long box. In both examples, the algorithm assumes the person will be standing upright and the dimensions will not be skewed by any other objects the person is holding, such as the box.
Such challenges with accuracy have been one of many issues plaguing the reputation of analytics for years in the video security market.
Advances using Deep Learning
Using Deep Learning, the program is fed many example images, and told those images represent a person, a car, an elderly woman, or any variety of very specific categories of objects the program may be tasked to classify unknown objects to. The major advancement of Deep Learning is that it is the software that determines what attributes are used for classification, and not the human developers.
The example images could number in the tens of thousands or greater, and the images may demonstrate the object from different angles, different light conditions, different regions of the world, and so forth. Because Deep Learning allows the software to determine object attributes based on real image examples, there are no preconceived notions as to what defines an object. Provided the image library fed to the program is sufficiently diverse there should be no inherent biases as to what attributes may define an object and no significant limit to the number of attributes which can be used for classification.
What the Future Holds
Deep Learning is still a relatively new technology; however, some say this technique may lead to computers being able to recognize objects better than people can, and with less data, in the future.
Presently, object classification is limited to how much training the VCA program receives, the diversity of the examples used to train the program, and the processing power available to perform accurate object detection and classification on video in real time.
Near term advances in algorithm training will come from developers using video instead of static images in the training process. Software trained using video clips could lead to VCA making classifications based on multi-faceted attributes. VCA could observe and note that cars travel on roads, whereas people walk on sidewalks. Attributes such as speed, movement patterns, where an object is located in the scene, walking gait, and other factors could be considered by analytics for better detection.
Training, detection improvements, and greater processing power combined with Deep Learning techniques could make near perfect accuracy a future reality for VCA.
About the Author
As Director of Product Strategy for Salient Systems, Brian manages Salient’s CompleteView Video Management Software and CompleteView Cloud VSaaS product lines. Brian has over 13 years of experience working with network cameras and video management products. Prior to his current position, Brian worked for Axis Communications as ADP Program Manager, Technical Trainer & Sr. Sales Engineer.