Objective To estimate the flexion and extension torques of the hip, knee, and ankle joints during straight-line walking using depth cameras and neural networks. Methods Gait information was collected from 20 individuals using an optical motion capture system, force plates, and an Azure Kinect depth camera. The subjects were asked to walk straight at their preferred speed while stepping on the force plates. The joint torques were obtained using visual 3D simulation as a reference value, and an artificial neural network (ANN) and long short-term memory (LSTM) network were trained to estimate the joint torques. Results The relative root mean square errors (rRMSEs) of the ANN model for estimating the joint torques of hip, knee, and ankle were 15.87%－17.32%, 18.36%－25.34%, and 14.11%－16.82%, respectively, and the correlation coefficients were 0.81–0.85, 0.69–0.74 and 0.76–0.82, respectively. The LSTM model had a better estimation effect, with rRMSEs of 8.53%－12.18%, 14.32%－18.78%, and 6.51%－11.83%, and correlation coefficients of 0.89－0.95, 0.85－0.91 and 0.90－0.97, respectively. Conclusions This study confirms The feasibility of using a depth camera and neural network for noncontact estimation of lower limb joint torques, and LSTM has a better performance. Compared with existing studies, the joint torque estimation results have better accuracy, and the potential application scenarios include telemedicine, personalized rehabilitation program development, and orthosis-assisted design.