Abstract —The pedestrian inertial navigation systems are generally based on Pedestrian Dead Reckoning (PDR) algorithm. Considering the physiological characteristics of pedestrian movement, we use the cyclical characteristics and statistics of acceleration waveform and features which are associated with the walking speed to estimate the stride length. Due to the randomness of the pedestrian hand-held habit, the step events cannot always be detected by using the periods of zero velocity updates (ZUPTs). Furthermore, the signal patterns of the sensor could differ significantly depending on the carrying modes and the user’s hand motion. Hence, the step detection and the associated adaptive step length model using a handheld device equipped with accelerometer is required to obtain high-accurate measurements. To achieve this goal, a compositional algorithm of empirical formula and Back-propagation neural network by using handheld devices is proposed to estimate the step length, as well as achieve the accuracy of step detection higher than 98%. Furthermore, the proposed joint step detection and adaptive step length estimation algorithm can help much in the development of Pedestrian Navigation Devices (PNDs) based on the handheld inertial sensors.