In order to ensure the accuracy of motion in complex terrains for quadruped robots, a real-time gait monitoring method based on machine learning and binocular vision was proposed. The residual visual error was addressed by identifying the optimal observation point for the foot end along the known gait trajectory, which was referred to as "capturing stillness amid motion". To compensate for measurement errors in foot position and posture due to inaccuracies in the vision system, an accurate prediction method based on a depth neural network was utilized. The simulation results indicate that the designed neural network achieves a 99.68% probability of attaining 0.025 mm foot position prediction accuracy, thus fulfilling the requirements for real-time and high-precision monitoring. It is concluded that the proposed method combines the generalization capabilities of machine learning with the complex error sources of the vision system to achieve high-precision measurement. Consequently, it provides a novel perspective for real-time and accurate gait monitoring of foot robots and serves as a valuable reference for foot end positioning and periodic gait maintenance.