Gait analysis is a powerful tool in the diagnosis and treatment of patients with neuromotor dysfunctions. Three-dimensional (3D), marker-based motion capture systems are hardly used in clinical practice, since they require long execution time, and technical experience. Markerless video-based tracking systems could represent a viable alternative to 3D marker-based apparatuses. The purpose of this study was to assess the performance of a low-cost markerless system, consisting of the open-source library OpenPose (OP), two webcams and a linear triangulation algorithm. Two synchronized videos of a healthy subject were acquired, with two webcams, in a walking session on a treadmill at comfortable speed. The lower limbs angular kinematics was simultaneously recorded with inertial sensors. 2-dimensional joint centers coordinates were assessed by OP, and computed in 3D by triangulation. The resulting sagittal joints angles were compared between the two systems, in order to evaluate the kinematic accuracy. The ability of OP in the detection of anatomical landmarks was analyzed in terms of reprojection error. Additionally, OP accuracy in measuring segment lengths was assessed. Results showed a mismatch between the points provided by OP on the two image planes, which led to reprojection errors up to 3.2 cm and segment length errors up to 3.4 cm. The angular traces, provided by OP, instead, were similar to the ones computed by IMUs, but OP inaccuracy affected mostly the computation of maxima and minima joint angles producing kinematic errors up to 9.1°.