This study proposes a method to enhance the accuracy of vision-based human joint angle estimation by correcting errors that occur during the process of estimating the positions and angles of human joints from videos captured with a monocular camera using long short-term memory (LSTM). The proposed method assumes joint angle data collected using a Vicon motion capture system as ground truth and trains an LSTM model to adjust the camera-based joint angle estimates to match the Vicon values. The test result demonstrates the potential of upgrading the performance of low-cost camera-based joint angle estimation methods to a level comparable with high-cost motion capture systems.