Object detection with Raspberry Pi3 and Movidius Neural Network Stick

Abstract

Object detection and classification is an increasingly important field of research in machine learning. Currently, powerful GPUs (Graphics Processing Units) are used to perform the computation-intensive operations in the shortest possible computing time. However, these systems are associated with high costs. In this paper a system for object detection and classification is developed, which gets by with less resources. This should minimize the costs while keeping the performance acceptable for the target application. To keep the costs low, a Raspberry Pi3 is used as development platform in connection with a Movidius stick for the outsourcing of the ANN. After explaining the theoretical basics of object detection and ANNs, this paper shows the implementation process of the selected hardware and software. For the evaluation of this system the algorithms YOLO and MobileNet are used and pre-trained models are used as basis. Based on the MSCOCO data set, both the quality of the object classification and the computing time are evaluated.