A Parallel Network Architecture for Automatic Slip Detection Task in Human-Centered Footwear Test

Abstract

Slips and falls are significant public health concerns, making slip-resistant footwear an essential component of prevention. This study aimed to develop an automated slip detection algorithm, designed to replace human observers in human-centered footwear testing protocols and provide a proof of concept for future real-world assessment. A parallel network architecture was proposed to classify videos into slip and no slip events, consisting of two distinct models: Inflated 3D convolutional neural networks (I3D) and Multiscale Vision Transformer (MViT). The two models operate in parallel, independently processing the input video data to capture diverse temporal and spatial features. To optimize the network, predictions from each model are aggregated by fusing their losses using a weighted averaging mechanism, ensuring a balanced contribution during training. The proposed approach was evaluated using two cross-validation techniques: 5-fold and Leave-One-Subject-Out (LOSO) to assess both overall performance and generalizability across subjects. Our proposed parallel network achieved the highest performance in both cross-validation techniques, with an accuracy of 94.63% ± 1.39% (5-fold) and 93.37% ± 3.02% (LOSO), and an F1 score of 94.33% ± 1.46% (5-fold) and 93.02% ± 2.93% (LOSO) on unseen test data. Notably, it outperformed standalone I3D and MViT models by approximately 7% and 4%, respectively.