The algorithm flow of an inertial-based Pedestrian Navigation System (PNS) can be divided into a trajectory-generation stage and trajectory-calibration stage. The Zero-velocity UPdaTe (ZUPT)-aided Extended Kalman Filter (EKF) algorithm is commonly used to resolve the trajectory of a walking person, but it still suffers from long-term drift. Many methods have been developed to suppress these drifts and thus to calibrate the trajectory generated by the previous stage. However, these methods have certain requirements, such as explicit map information or frequent location revisits, which are hard to satisfy in such situations as Search and Rescue (SAR) operations. A new approach is proposed in this paper that requires no explicit presupposition. This approach is based on a particle filter framework, with the weight of particles being adaptively adjusted according to the a priori knowledge of building structures and human behaviours. The distribution of particle weights is designed with awareness of the regular structures of buildings. The time-varying parameter of the distribution is acquired from a Hidden Markov Model (HMM) based on the foregoing odometry, which has a close relation with human behaviour. HMM is trained offline based on samples acquired in advance. Many real-world experiments under various scenarios were performed, and the results indicate good accuracy and robustness of the proposed approach.