In this paper, a real-time implementation of a second-order adaptive windowing technique for acceleration estimation is presented for a slider crank mechanism. It is shown that there are some critical parameters which should be carefully set to proper values for a satisfactory acceleration estimation. In this technique, the size of the window is automatically enlarged until the position errors will not be out of the error bands defined by the user. In order to show the effectiveness of the second-order adaptive windowing technique in real-time, a test setup composed of a slider-crank mechanism is used. In the performed tests, only the crank angle is measured via an incremental encoder having a 1024 pulse per revolution, and then, the slider acceleration is estimated by the help of some kinematic calculations using the outputs of the second-order adaptive windowing technique. The estimated acceleration results are directly compared with an acceleration sensor having a measurement range with ±6g with a nominal 0-200 Hz bandwidth attached on the slider. It is found that the root-mean-square value of the acceleration errors along the entire motion profile is at tolerable level so that this effective acceleration estimation technique could be applied to various on-line robotic applications as a soft sensor.