This thesis set out to find the answer to the following research hypotheses: it is possible to determine whether a person is taking something from a shelf in a real life retail environment with a smartphone that is located in the person's pocket, and whether that user returns the object to the shelf or not, and that it is possible to achieve similar classification accuracy regardless of whether the used classifier is trained with the same user's activity data or other user's activity data.
Four users carried smartphones in their pockets while shopping in a real life retail environment over 12 occasions, and the sensors results of the phone's accelerometer were recorded and manually labeled to one of four common shopping activities, including taking a product from the shelf for inspection, and taking the product to a basket or trolley. The raw data was copied to a computer and forced to equal time frames. L2 norms, basic statistical features and Fourier transform were extracted from the data as features. These features were then applied to K-nearest Neighbors and Random Forest algorithms for classification.
The methodology was evaluated considering overall classification accuracy, the effect of window size on classifier performance, and the effect of training the classifier with only the user's data it is classifying versus training it with several user's data. The results of the evaluation indicated that the chosen methodology did manage to classify the shopping activities to some extent, but not well enough to comphrensively conclude whether as user is taking something from the shelf, and whether that user is returning it to the shelf, revoking the first research hypothesis. The evaluation also indicated similar classification performance in using one user's data for training versus using several user's data, supporting the second research hypothesis.
ACM Computing Classification System (CCS):
Hardware
Communication hardware, interfaces and storage
Signal processing systems
