Early fall detection is an important issue during gait rehabilitation training. This paper proposes an approach for pre-impact fall detection during gait rehabilitation training based on a 3D convolutional neural network (CNN). Firstly, pre-training data is collected and used to pre-train the 3D CNN to differentiate between a normal walking and a fall based on their general spatio-temporal patterns. Secondly, fine-tuning data is created by combining the pre-training data with a 3second normal walking sample collected from a new trainee whose falls are to be detected. The pre-trained 3D CNN is further fine-tuned by the fine-tuning data to learn the spatiotemporal patterns of the new trainee. Finally, a temporal sliding window is used to feed video snippets into the fine-tuned 3D CNN for fall detection. To the best of our knowledge, this is the first pre-impact fall detection approach based on a 3D CNN using RGB images. Moreover, the training strategy used to train the 3D CNN can alleviate the generalization issue of the 3D CNN when only limited training data is available in gait rehabilitation training. With 225 testing trials from 5 trainees, the proposed pre-impact fall detection approach achieves a detection accuracy of 100% within 0.5 second after falls start. Experiment results show that this approach is efficient, accurate, and practical in achieving intelligent fall detection during gait rehabilitation training.