Sensor-based maximum-likelihood estimation of item assignments
In a next-generation retail store, the presence of a shopper at a particular shelf location, at about the time that an item was removed from that shelf location, gives rise to a hypothesis that the item may have been picked for purchase by that shopper. An ensemble of different such hypotheses, about different items removed by different shoppers, is evaluated jointly to determine which hypotheses are most likely. Other factors can influence the determinations, such as a physical distance between picked products, together with a time interval between their picking. If cameras are used to detect removal of items from shelves, hand depictions captured by such cameras are further factors that may be considered, e.g., helping associate certain item removals with a common shopper. By generating purchase lists for each shopper with such methods, shoppers can be relieved of the need of having each item scanned at a checkout. Some embodiments are privacy-preserving—neither employing nor capturing personally-identifying information. Other embodiments don't require tracking movements of shoppers through the store. A great number of other features and arrangements are also detailed.
1. An image processing system comprising:
a first set of one or more cameras associated with a first rack of shelving, and in communication with one or more processors that are adapted to recognize items on said first rack of shelving using a first convolutional neural network that is configured to recognize a first set of items;
a second set of one or more cameras associated with a second rack of shelving, and in communication with one or more processors that are adapted to recognize items on said second rack of shelving using a second convolutional neural network that is configured to recognize a second set of items;
wherein the first rack of shelving is adjacent to the second rack of shelving, and the first set of items is different than the second set of items, and wherein the second convolutional neural network is trained to recognize a particular item that the first convolutional neural network is not trained to recognize, and the first convolutional neural network is trained to recognize a certain item that the second convolutional neural network is not trained to recognize, and wherein the first set of items comprises a serialized steganographic watermark marking.
2. The image processing system of claim 1 further comprising a neural network configured to determine feature data associated with a hand or arm attribute in a field of view of the first set of one or more cameras.
3. The image processing system of claim 2 in which the feature data is provided to a neural network as an input to determine a gender associated with the hand or arm attribute.
4. The image processing system of claim 1 further comprising a neural network configured to determine feature data associated with a hand or arm attribute in a field of view of the second set of one or more cameras.
5. The image processing system of claim 4 in which the feature data is provided to a neural network as an input to determine a gender associated with the hand or arm attribute.
6. The image processing system of claim 1 in which each of the second set of items comprises a serialized steganographic watermark marking.