An image recognition function is indispensable to detect objects in an image for human detection systems that support the safety of industrial vehicles. Image recognition using deep learning requires training with a large number of images of the detection target in the actual operational environment, and the process of acquiring images required a great deal of work. Therefore, in collaboration with the National Institute of Advanced Industrial Science and Technology, Mitsubishi Heavy Industries, Ltd. has developed a multi-step training method that augments the training data using public data and computer graphics which are close to the conditions of the actual operational environment and allows efficient training with these data in steps. It has been confirmed that this method can achieve the same detection performance with only 280 actual images as that achieved by the conventional method with approximately 5,000 actual images. We plan to widely apply the developed technology to our products that have image recognition functions using deep learning.