An Ensemble of ConvTransformer Networks for the Sussex-Huawei Locomotion-Transportation (SHL) Recognition Challenge

Abstract

The task of the SHL recognition challenge 2021 is to recognize eight modes of locomotion-transportation based on radio sequential data collected by smartphones. These data includes GPS reception, GPS location, WiFi reception and GSM cell tower scans. In this challenge, our team (Transformers) presents a recognition scheme. First, a deep model (ConvTransformer) composed of convolutional and Transformer subnets is proposed. The convolutional subnet captures local features, and the Transformer subnet constructs longterm dependencies. Then, the ConvTransformer network is used to recognize different locomotion and transportation modes through location data. Finally, since the still and subway categories are easily confused, our team uses another ConvTransformer network to further classify them through cells and location data. Through cross validation techniques on the training dataset, our proposed scheme achieved a F1 score of 0.6838 on the validation dataset. CCS CONCEPTS• Human-centered computing -> Ubiquitous and mobile computing; • Computing methodologies -> Neural networks;

Accuracy Improvement of Vehicle Recognition by Using Smart Device Sensors

“…Ref. [38] distinguished between users standing still, walking, running, cycling, driving a car, or riding a bus, subway, or train to yield a best F1-score of 0.8779. The use of GPS sensors in [36][37][38] provides less specific displacement sensitivity than the gyroscope and accelerometer used in our study, a difference that can at least partly explain our higher accuracy of 98%.…”