Development of an algorithm for detecting moving objects in images from a real-time video stream

Abstract

One of the tasks of recognizing objects in video is to find their location in the current frame. In algorithms that use neural networks, such as SSD[1], YOLO[2], search and classification are performed by a single model and are inseparable steps. In the case where there is a computing node in the system that has enough power for complex image processing and can do it before the server that recognizes objects, separating the object search and recognition operations can reduce the latency, load on the network and the recognition node.