Indoor localization is the basis for most Location-Based Services (LBS), including consumptions, health care, public security, and augmented reality. Sensory landmarks related to the indoor spatial structures (such as escalators, stairs, and corners) do not rely on active signal transmitting devices and have fixed positions, which can be used as the absolute positioning information to improve the performance of indoor localization effectively without extra cost. Specific motion patterns are presented when users pass these architectural structures, which can be captured by mobile built-in sensors, including accelerometers, gyroscopes, and magnetometers, to achieve the recognition of structure-related sensory landmarks. Therefore, the recognition of these landmarks can draw on the mature methods of Human Activity Recognition (HAR) with improvements. To this end, we improved a Long Short-Term Memory (LSTM) neural network to recognize different kinds of spatial structure-related sensory landmarks. Labels of structural sensory landmarks were proposed, and data processing methods (including interpolation, filter, and window length) were used and compared to achieve the highest recognition accuracy of 99.6%.