This study explores a model for detecting fall, non-fall and near-fall events as frequent experiences of near-falls are closely associated with a heightened risk of falls. Detecting near-falls can lead to more accurate predictions of falls. However, near-falls exhibit certain movement patterns similar to actual falls, making it challenging to distinguish between the two. We investigate the detection of fall-related activities, including falls, near-falls, and non-falls, by utilizing dynamic motion images derived from video clips. There are two primary approaches for classification: a vanilla convolutional neural network (CNN) model and a transfer learning approach that utilizes InceptionV3 and DenseNet201 models as feature extractors and train conventional machine learning classifiers, such as support vector machine (SVM), K-nearest neighborhood, decision tree, and random forest, and adaptive boosting models. The vanilla CNN model achieved a high accuracy of 97.89% compared to the transfer learning approach, which reached a maximum accuracy of 95.31 for binary classification of fall and non-fall events. On the other hand, the transfer learning approach, which integrated feature from InceptionV3 and DenseNet201 into machine learning classifiers, achieved an accuracy of up to 90.14% for the three-class classification of fall, non-fall, and near-fall events. This underscores the model’s robustness in detecting various fall-related activities, highlighting its potential for improving safety in at-risk populations.