Location Awareness and Location-Based Services facilitate our daily lives by involving in broad domain of applications, including commercial, healthcare, and life-saving services. Global Navigation Satellite Systems (GNSS) provide satisfactory outdoor localization services, in contrast, indoors, GNSS performance degrades significantly due to signal blockage and attenuation. Many Indoor Localization Systems (ILS) have been proposed aiming to improve the indoor localization performance. Pedestrian Dead Reckoning (PDR) using off-the-shelf inertial sensors has gained great popularity as low-cost autonomous technique. PDR provides a reliable short-term solution bridging outages of wireless techniques. Contrarily, its performance degrades rapidly due to the inherent errors of low-cost inertial sensors and the accumulated drift of the gyroscope heading over long period of time. Despite numerous research efforts, the heading problem and the unconstrained smartphones cases still limit the widespread use of PDR. In this paper, we propose robust PDR scheme to achieve better performance for long period navigation under different smartphone poses. To improve the length estimation, we propose a new step detection method based on a robust peak and valley algorithm and a nonlinear model to estimate a robust empirical length regardless smartphone pose. The magnetic field quasi-static periods are used to accurately calibrate the gyro heading. The pedestrian walking between two anchor nodes is exploited to further calibrate the step length and heading. Finally, a real-time PDR outlier filtering is proposed based on robust turn detection and Random Sample Consensus (RANSAC) algorithm to curb the PDR from the slight body shaking. Extensive experiments are conducted to evaluate each PDR component. Our results show that reliable step counting is achieved with different smartphone poses, additionally, the step length is estimated with a maximum error of less than 2% of the walked distance. Moreover, the estimated heading after 20 minutes walking drifted less than 5º. The overall performance of PDR is significantly improved in long-term scenarios with a mean and a 90% error of 1.83 and 4.1 m, respectively, over the entire 750-m continuously walking distance. Furthermore, when PDR is combined with Bluetooth Low Energy (BLE) beacons for step length, heading, and position calibration, the corresponding values are improved to 1.45 and 2.35 m, respectively.