Landmark Detection Based on Human Activity Recognition for Automatic Floor Plan Construction

Landmark detection technology has a wide range of applications in people's lives, including map correcting, localization and navigation, etc. Besides, landmarks are also utilized to label different areas for automatic floor plan construction. Currently, vision-based landmark detection methods have some limitations, such as light, camera shaking, and privacy-invasive. In addition, deep learning-based methods increase the time consumption of marking labels due to the huge requirement for data. Targeting the above challenges, our work first proposes a landmark detection approach based on Human Activity Recognition (HAR) for automatic floor plan construction, which introduces a self-attention model to recognize various landmarks by walker's daily activities due to their strong correlation. First, the accelerometer and gyroscope sensor data are extracted and eliminated by a Gaussian filter and are divided into the same length segments by slide window. Next, it is input into the self-attention network to train a human activity recognition model. Finally, the corresponding relationship between human activities and landmarks is created to detect landmarks through the trained HAR model. Empirical results on two publicly available USC-HAD and OPPORTUNITY datasets show our proposed approach can recognize landmarks effectively.