Accurate user intent recognition is vital to the success of achieving volitional control of rehabilitation robotics. Real-time ultrasound (US) imaging of skeletal muscle, or sonomyography, is an alternative noninvasive sensing mechanism for device control. The objective of this study was to evaluate sonomyography for continuous estimation of hip, knee and ankle-joint moments during multiple ambulation tasks. Ten able-bodied subjects completed level, incline and decline walking while equipped with a portable US transducer on their anterior thigh. Multiple time-intensity features were extracted from US images of the knee extensor muscles collected during the three ambulation tasks. Hip, knee and ankle moments were continuously estimated by Gaussian process regression models in both fully subject-dependent and partially subject independent frameworks. A two-way analysis of variance was completed to assess the effect of subject independence as well as joint level (hip/knee/ankle) on the moment estimation. Subject Dependent regression models resulted in the lowest error for estimation of hip, knee and ankle moment during all three ambulation tasks in comparison to partially subject-independent regression models (p<0.01). Remarkably, within the subject dependent regression models there was no significant difference in the mean error of moment estimation when comparing across the three joints, with mean percent errors as low as 0.74%, 0.68%, and 3.02% for the hip, knee, and ankle, respectively. Despite only capturing sonomyographic features from the anterior thigh, this high-dimensional sensing data can be used to accurately estimate changes in both proximal and distal joint kinetics during varying ambulation tasks.