A deep neural network (DNN)-based Wi-Fi/pedestrian dead reckoning (PDR) indoor positioning system using an adaptive robust factor-graph model is proposed in this study for the indoor positioning of smartphones. In Wi-Fi positioning, the authors use a DNN to extract robust features from fluctuant Wi-Fi signals in the off-line phase, and obtain more accurate positioning results by computing posterior probabilities in online positioning. Acceleration, gyroscope, and magnetometer data are used to calculate attitude angle, step frequency, and step length, respectively. Received Wi-Fi signal strength is susceptible in complex indoor environments, and PDR errors accumulate over time. A factor-graph model with adaptive robust adjustment is proposed to fuse the positioning results of Wi-Fi and PDR, and it overcomes such shortcomings as slow update frequency and gross errors of Wi-Fi and PDR errors accumulated over time, respectively. When the absence of PDR occurs, hidden Markov model is introduced to smooth multiple DNN-based Wi-Fi positioning estimates at the unknown point to obtain the optimal solution. Experimental results show that the proposed system is more robust and has better accuracy under different motion gestures (held-in-hand, dangling, and calling).