In order to solve the problem of high cost and low efficiency of robot arm calibration, the low-cost and real-time calibration method was investigated. A low-cost binocular camera was selected as the measuring equipment， based on the differential transformation theory, the establishment method of link parameter error identification model was introduced. The measurement configuration pool was generated by inverse kinematics algorithm with the work trajectory and marker visibility as constraints. On this basis, the measurement configurations were optimized by DETMAX algorithm to establish the identification model, and a matrix balance method was also proposed to make the model better-conditioned. At the same time, K-means algorithm was used to cluster the selected configuration, which were used as multi-group intermediate configurations transferred to the control system to generate several smooth work-calibration trajectories. Finally, the identification accuracy of link parameter errors under measurement noise with different intensities was verified by simulation. The results show that the matrix balance method increases the observation index of the identification model from 10.6 to 6.2×104, and the condition number is decreased from 1.2×103 to 37.8, which improves the performance level of the model significantly. The well-conditioned model makes most of the DH error identification results insusceptible to the measurement noise.The standard deviation of measurement noise is increased from 0.1/3 mm to 1/3 mm, the average identification deviation of link parameter errors only increase from 1.1% to 1.4%. Therefore, the new method can meet the low-cost, real-time and high-precision calibration requirements of the robot arm.