Indoor Non-Visual Semantic Landmark Recognition Based on the Multi-Channel CNN-Transformer Network

With the rapid development of smart devices and sensor technologies, time-series data based on Inertial Measurement Units (IMU) signals have been widely applied in various fields, particularly in scenarios such as autonomous driving and indoor positioning. However, existing traditional methods often struggle to effectively process and extract valuable features from data collected by different types of sensors in complex environments, which limits classification accuracy and robustness. This paper proposes a Multi-Channel CNN-Transformer (MCCT) network to classify IMU time-series data. The network extracts features from accelerometer and gyroscope signals through parallel convolutional branches, and combines channel attention recalibration with multi-layer Transformer encoder to efficiently classify IMU time-series data. The performance of the model is evaluated through key metrics including accuracy, precision-recall (PR) curves, and confusion matrices. Experimental results show that MCCT performed excellently across multiple datasets, achieving high accuracy and good robustness in different sensor data scenarios.