The distribution of ground reaction forces (GRFs) on the soles of a legged robot changes the location of center of pressure (CoP) or zero moment point (ZMP) during walking, which is crucial for maintaining dynamic stability of legged locomotion. In addition, many lower-limb exoskeletons use GRF to detect different phases in a gait cycle as well as the walking intention of the wearer. Therefore, many legged robots and exoskeletons install force or pressure sensors on the feet to measure GRF. However, frequent contact with the ground tends to damage the sensors on the feet, reducing reliability and durability of the robot. In this paper, we propose an on-line GRF estimation method based on knee joint acceleration. We analyze the experimental data and identify features in the knee joint acceleration that are highly correlated to GRF. A linear dynamic model is used to fit experimental data and generate GRF estimates. Then by searching for the desired features in the knee joint acceleration and applying the linear dynamic model, we can online estimate GRF with reasonable accuracy. Consequently, force or pressure sensors on the feet of a robot are no longer required, and reliability of the robotic system can be improved.