Aiming at the issues of real-time detection and positioning of movement target, a method of deep learning is proposed. The Single Shot multibox Detector (SSD) detection method is used under Caffe framework, the VGG16 model is used as the basic network model, and the additional feature convolution layers are used to extract the multi-scale convolution features. Then the experimental data set is iteratively trained to get the motion target detection model. The moving objects are detected using the trained model and then positioned through binocular vision positioning method. The experiment results show that the system can reach 40 fps. In the 10 m×10 m scene, the average positioning error is within 6 cm. The system has sound performance both in speed and precision, which provides an effective solution for the real-time detection and positioning of human motion in large and medium-sized scenes.