The current unmanned retail business is booming, all kinds of vending machines in the original function of continuous integration with the new retail mode, and the current cosmetics vending machine function is relatively single, in order to meet the personalized needs of customers, the need to add new functions on the basis of the original function of the vending machine. According to the current situation, this work takes different cosmetics recommended by different skin color groups as the improvement direction, and uses raspberry pie and peripherals combined with visual processing technology to build a set of cosmetics sales system based on the Internet of Things. The design content of this work is arranged as follows:(1) Propose the overall plan according to the commodity sales process, and complete the preliminary design of the system hardware circuit and software algorithm according to the plan;(2) The hardware system is designed for the research plan, and the selection is carried out according to the order of central processing, human-computer interaction, visual acquisition, signal transmission and shipment, and at the same time, the hardware of each part of the system is integrated, vending machine hardware platform to build;(3) the use of C combined with QtCreator to complete the front-end interface development, design two sales mode: optional and recommended, and for different sales mode to complete the shopping process platform to build. On the basis of the completion of the overall design, this work also focuses on the recommended model. First, Python combined with OpenCV is used to realize face detection through Haar cascade classifier. After detecting face data, the last frame of detection image is saved. Secondly, for the saved face image, Haar is used to divide the facial features of the face again, and the appropriate area is collected as the skin color analysis object. Finally, skin color analysis and commodity recommendation are completed by traversing RGB pixel points of the divided image and mapping it to YUV space. Through the system debugging and simulation, this work successfully realizes the two sales modes of self-selection and recommendation, and verifies the feasibility and reliability of the design scheme and algorithm through the test results. The design for the vending machine sales model for innovation and optimization, for its future development direction provides a novel idea.