Foot scanning and shoe size recommendation using a low-cost Kinect

Abstract

With the development of the Internet and transportation, online shopping has increased rapidly. However, the customer experience has not been satisfactory because they often experience mismatches when choosing the correct size online, especially for footwear. In general, shoe size selection is usually based on shoe fitting, which involves foot dimensions or past customer experience. While structure- or laser-based scanners are the current gold standard for 3D modeling and measurement, their high cost and size limitations constraint their use in the customer’s home. The primary goal of this research is to design and validate a novel 3D foot scanning and shoe size recommendation, utilizing a low-cost, portable Kinect device for practical use in home applications. The scanning accuracy of the Kinect based prototype foot scanner was investigated in two steps. Static objects, including a standard cylinder and mannequin foot were used for testing in the first step. The accuracy was estimated by comparing the Kinect scans in the form of point clouds with those obtained with a high-resolution structure light-based scanner (as golden reference). The experiment resulted in deviations of 0.99±0.70mm and 0.97±0.86mm, for the cylinder and mannequin foot, respectively. Real human feet were tested in the second step, and the experimental results showed that the average deviation was 1.14±0.85mm, and the absolute error of 95% points was within 3.25mm. Therefore, a low-cost, portable Kinect device can be used to obtain 3D foot scans for different design and anthropometric applications. Simulated measurements (SM) were developed in this study, including alignment of 3d canned food model and foot measurement extraction algorithms. To further validate the developed method, comparing the SM extracted from the ideal foot models (complete foot models scanned from a golden reference scanner and a commercial platform) with manual measurements (MM as golden reference). Paired t-test results showed foot width, short heel girth, mid-foot height and ankle girth dimensions significantly different between the SM and MM methods. However, intraclass correlation coefficient (ICC) results for all dimensions showed good to excellent consistency（ICCs>0.85）and the mean absolute percentage error (MAPE) within 2% except for girth dimensions. To further validate the Kinect sensor as a device for measuring foot dimensions, this study evaluated the SM performance on the Kinect scanned foot model. For the SM extraction performance on the Kinect, foot length (ICC=0.92) and short heel girth (ICC=0.85) dimensions demonstrated acceptable agreement with reference measurements (MM) and their MAPE, all within 2%. A fitting shoe experiment was conducted to check the feasibility of using a low-cost portable Kinect as a shoe size recommendation device. A linear regression-based shoe size recommendation using Kinect extracted measurements was built and validated. The experimental results showed that the Kinect-based linear regression shoe size recommendation model could reach the exact shoe size accuracy of 50.00%. The accuracy rate for within one size difference was 86.36%. This level of accuracy could help the customer narrow the range of the optimal shoe sizes for online shopping. This study firstly validated the single Kinect-based shoe size recommendation and thoroughly validated the Kinect scanning accuracy. In addition, the Kinect-based foot dimension extraction algorithm was developed and verified in this study. The findings from this study indicated that a low-cost, portable Kinect scanning device could provide an alternative option for foot scanning and shoe size recommendation.