Decoupled Spatial-Temporal Representation Learning and Gradient Boosting for Multi-Axial Locomotion Tracking

Human Activity Recognition (HAR) using tri-axial accelerometer data remains a fundamental challenge due to high-frequency noise, temporal dependencies, and device placement variances. This paper introduces the machine learning pipeline developed by Alone Team for the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge. Utilizing an early data-level fusion schema, raw three-axis acceleration streams are segmented into 500-timestep windows and stacked into a unified multi-dimensional tensor. A Deep 1D Convolutional Neural Network (1D-CNN) is trained to automatically extract spatial-temporal embeddings, compressing complex structural waveforms into a dense 128-dimensional bottleneck vector. These frozen representations are then evaluated by a downstream LightGBM classifier optimized via leaf-wise tree growth to isolate robust multi-class cluster boundaries. The structural variance across unseen device positions introduces a challenging generalization ceiling, yet the framework maintains consistent architectural stability across divergent locomotion profiles. Finally,a sample-by-sample broadcasting post-processing schema was implemented to format predictions into a strict 92726 × 500 matrix for standardized evaluation.