Falls in older adults is a major public health issue with approximately one-third of individuals aged 65 or older experiencing at least one fall event annually. Accordingly, there is a need for methods to identify elderly individuals at risk of falls. Methods allowing automated and instantaneous assessment of fall risk would have considerable utility in hospitals and nursing care facilities with staff shortage or time pressure. The present study evaluated models for estimating fall risk from gait characteristics measured during a single gait cycle. As gait images are affected by clothing, skeletal data was recorded using motion capture imaging. Fall risk was determined according to the history of falls within the preceding year. Of 80 healthy subjects aged over 65 years who participated in this study, 45 had experienced falls in the preceding year. Gait features, time series data, and gait energy images were recorded. The area under the receiver operating characteristic curve (AUC) was utilized as a performance measure to evaluate machine learning models. The input of Gait Energy Image data into a 6-layer convolutional neural network (CNN) provided higher accuracy (AUC = 0.67) than other inputs. Visual explanations from the 6-layer CNN created using Eigen-CAM demonstrated that areas associated with step length were predictor for estimating fall risk. Arm swing at heel strike and feet movements were also predictors of fall risk. The models evaluated in this study can be utilized to estimate fall risk from instantaneous measurements, with promising applications in various industries including medical care.