At present, it is not using machine vision technology to identify commodity trends, but to use radio frequency (Radio Frequency IDentification, RFID) technology, but in fact, all developers know that RFID technology can not move to the end Of the unmanned store is ideal, but in the current concept of the beginning of the time to help people show the concept of unmanned store concept transition technology to be able to achieve Amazon's "Just Walk Out" unmanned store experience, still need to Machine vision technology is completed.
However, the current machine vision also has its technical limitations, but it can be overcome. This article will detail the limitations of RFID technology in unmanned stores, and how to use machine vision technology to explore the "new retail", "smart retail" The possibility of.
Just Walk Out is still a test
Amazon Go's important concept is "Just Walk Out", the consumer into the store, as long as the shelves from the want to buy the goods, without any other checkout procedures can be left to complete the identity of the store, and use Amazon Go has not been able to complete this ideal, the most important reason is that Amazon Go directly to the target as a large unmanned shopping malls, but so far whether using RFID technology or machine vision, are unable to complete In the case of a large number of unmanned stores, it is possible to allow only one consumer to enter the market at once. In terms of technological development, it is important to really achieve Amazon Go. ~ 3 years of development time.
Unlimited store is the biggest limitation is that the store can only accommodate a guest at one time, the system can not determine a number of guests, more shopping cart order consumption patterns. More than a single is a bottleneck to be broken, but also the next technology development Important goals. At present, our team has been in the development of more than a single machine vision checkout technology, in the near future will be able to see the experimental shop show, but before that, we choose to kiosks, beverage cabinet mode first The implementation of intelligent retail, they save more than a single concern, so that consumer experience more in line with consumer expectations.
Consumer habits have not yet developed
In fact, the concept of "unmanned stores" is very unfriendly to the consumer, and technology should be a store that can be used to help humans to become more convenient and more intuitive, and may not be the way consumers are looking for So we can see that there are many unmanned stores in China failure cases, you can see the consumer in order to avoid the checkout and tear off and casually lost on the ground of the RFID label paper; or because of storage management inaccurate, the store unmanned maintenance , Left the house full of rotten food scrapped.
Therefore, not so much a "unmanned store", I think this wave of the trend should use the "smart retail" concept is more accurate, and the use of machine vision identification technology, to truly show the wisdom of retail intelligence.
The world's largest visual competition ImageNet in 2017 competition, the winner can do the machine vision accuracy rate is 81.7%, for commercial technology is still too low.Our solution is to use four cameras at different angles Shooting, using a video camera for one second, that is, one second to capture 120 photos, and then check each other to determine the action of consumers and the purchase of goods. At present, through this way to achieve 93% accuracy rate The
RFID high cost does not match retail applications
Unmanned stores are a retail concept that starts with an employer's position and hopes to reduce the cost of personnel. It is not a innovation that takes the interests of consumers as a starting point.However, RFID tags are very expensive to build, and a one-time RFID tag The cost of about NT $ 3 or so, and the cost of change, retailers will be difficult to afford.
Machine vision devices such as cameras and GPUs are fixed overhead and are a one-time expense. The only change is the establishment of the image model of the product, but still close to the fixed cost. When the store more goods, the more stores, the marginal cost of the more This is the hard device for retailers.
At present, our team has been able to set the camera vision accuracy to 93% through the establishment of multiple cameras, although 97% to 99% of the commercial standard is the level of passing, but if the retail business personnel The cost of about three percent, so that the accuracy rate of 7% down is still much lower than the cost of three people, machine vision in the application of intelligent retail is still worthy of development and profitable technology.