Adaptive treadmill is a research hotspot in rehabilitation medicine and ergonomics, and also an important part of virtual reality motion input devices. In this paper, aiming at the problem that the position of the human body on the adaptive treadmill is almost unchanged relative to ground and it is difficult to obtain the speed of the human body through simple position difference, a wide adaptive, unmarked and noncontact walking speed estimation method is proposed. Quaternion calibration, Gaussian filtering and cubic spline interpolation processing are performed on the position data of the human joint points collected with Kinect, the step length correction algorithm is used to calculate the spatiotemporal gait parameters during walking. The speed of the user walking on the adaptive treadmill is estimated based on the gait time and space parameters. The speed estimation value was compared with the actual treadmill speed set on a fixed speed treadmill, the result verifies the effectiveness of the speed estimation algorithm. The proposed speed estimation algorithm can be applied to the development of the control algorithms for adaptive treadmills.