A new type of robot collision detection algorithm was proposed for the security problem that collision may occur when conventional industrial robots operate in an unknown environment. The convolution torque observer was designed. The robot collision detection was realized by real-time observation of the deviation between the joint output torque and the dynamic estimation torque. The static LuGre model was used to compensate the joint friction in order to avoid the interference of joint friction of the robot in different poses and motion states on robot collision detection. By monitoring the motion of actual industrial robots, more accurate static LuGre model parameters were identified. The collision detection algorithm does not need acceleration information, avoiding the calculation error caused by the secondary derivation of the position feedback information. The joint torque was acquired based on the current information of the joint servo drive. It is not necessary to install a special force/torque sensor. Therefore, in the case of conventional industrial robots without additional configuration, just collect the robot joint drive motor current and position information to achieve collision detection. The effectiveness of the collision detection algorithm is verified by human-robot interaction experiments.