Artificial intelligence has become the security guard in retail. Neural networks are used to reduce losses, control compliance with regulations, increase customer loyalty – and this is not a complete list of tasks that machine vision can perform. Why do we need to use computer vision now and how to properly introduce new technologies in retail?
What is computer vision and what tasks does it solve?
Machine or computer vision is a part of artificial intelligence that, using neural networks, allows you to understand what is shown in a picture or video sequence. It helps to find, for example, forgotten things, a car or a person. In addition, the neural network can not only determine what kind of object is in the frame, but also its state: the barrier is open or closed, whether a fire has occurred. This allows you to expand the use of computer vision in security systems and introduce control functions in manufacturing, retail and other industries. Also, computer vision can in some way “predict” the future, for example, the neural network understands that a car has arrived at the warehouse and loading operations should begin.
If you describe the work of computer vision in simple terms, then in order to teach a neural network to read certain information, you need a special specialist – a data scientist who deals with data analytics. He has experience in developing neural networks and works with video images and pictures. Technically, this process is divided into several stages: 1. Data collection. A video sequence or images are selected that neural networks need to learn to distinguish from each other. 2. Setting up the neural network. The actual, suitable architecture of the neural network is selected. The choice depends on what accuracy, speed and what task it faces. 3.Neural network training. The neural network is provided with images or video, which it reads in a special way. 4. Network installation. The trained neural network is embedded in the customer’s infrastructure or in the cloud, where images are received that contain already familiar objects: their neural network can read.
What tasks does computer vision help solve in retail?
There are typical requests in retail – determination of “holes” on the shelves, crowded places in the mall to obtain objective data on the possible cost of renting advertising space… But more often these are unique needs for each customer.
Computer vision is primarily relevant for large networks. The tasks at the sites are very similar, so it is enough to train the neural network once, and then replicate it to all stores.
For example, one large network selling construction products ordered neuron training to identify visitors with large bags and backpacks at the entrance to the trading floor, because expensive tools are stolen from them in this way. For another FMCG chain, which is already working on the Scan & Go model, we are preparing a solution to identify visitors who took goods out of the store without payment. A solution based on a speech recognition system is already being developed for the gas station network to monitor the compliance of employees with regulations for communicating with customers.
In general, the neural network can be used for a wide range of tasks, but we recommend it only for digitizing important and repetitive processes that occur in each store and can, for example, cause serious damage: financial, reputational or security related. But what if, over time, new tasks appear or the objects that the neural network reads change, for example, the assortment of a store or the packaging of some product? – The neural network can be simply re-trained. Of course, it is cheaper to retrain the neural network in batches, for example, 100 new objects each, rather than retraining with each new one. After training, a new version is released and uploaded to a server or cloud.
For which customers is computer vision the most cost-effective?
According to DZOptics, computer vision is primarily relevant for large networks. The tasks at the sites are very similar, so it is enough to train the neural network once, and then replicate it to all stores. A trained neural network does not need to be serviced in any special way, and the camera and the server need standard maintenance. Usually, the introduction of computer vision takes about three months: the first version is made in a month, then it can be further trained in another 2 months. However, it all depends on the complexity of the tasks.
How popular is the use of computer vision in retail now?
They really began to talk a lot about this, interest is growing, but so far not everyone has decided to work with artificial intelligence. People want to buy a box, plug it into an outlet, and have it work. But with a neural network, this will not work, at least not yet.
First, to train the neural network, you need to collect a sufficient amount of material. It takes time, sometimes requires additional filming, which often stops the customer. People want to buy a box, plug it into an outlet, and have it work. But with a neural network, this will not work, at least not yet.
Secondly, often the decision to implement computer vision is at the mercy of IT departments, but they do not have expertise in reducing losses, increasing sales, combating shoplifters, etc. Only the security service and the operations department have it, which, in turn, do not understand the technical aspects of neural networks. That is why many companies have an innovation directorate that establishes communication between IT, security and other departments.
In fact, this is a good niche for the growth and development of SB employees. Their powers are now being eroded, and here there is a chance to transform towards security experts in order to closely interact with specialists who will train neural networks.
What is the future of computer vision in retail? Will neural networks be able to completely replace physical security?
They will partially take over their functions, help reduce the number of posts, and hence the cost of them. How does a security guard work? He observes the buyers, collects data, analyzes them and decides whether to check / detain someone or not, and then acts. All of these steps, except the last one, can be automated using computer vision. But the action itself – to detain the shoplifter – the neural network will not be able to do. Perhaps in the near future, observers will simply leave and only those guards who are able to act effectively will remain.