A New Approach to Use Modern Object Detection Methods More Efficiently on CCTV Systems

Abstract

DL architectures rely on extensive usage on powerful computer systems to operate in real-time. Therefore, cooperative and constructive optimizations should be made in both architecture and software parts of the related DL system. In this work, input system of the YOLO architecture is modified to accept several sources at the same time with two effective methods to increase the efficiency of the hardware system. First method is to design a scheduler which will allow YOLO architecture to process several input sources sequentially, allowing the architecture to use its full potential. Second method is to design a preprocessing algorithm to combine 4 or 9 input sources in a single input source as a 2x2 or 3x3 image matrix. In this way, YOLO architecture processes four or nine times more images in the same time, increasing its practical frame per second (FPS) value by four or nine folds. Experiment results on our machine show that the used YOLO architecture can process 3 input sources at the same time with only minimal loss of accuracy of 0.002 in terms of Mean Average Precision (mAP) while using the proposed scheduler. Additionally, using 4 inputs combined increases the practical FPS value from 31 to 120 and using 9 inputs increases the practical FPS value from 13 to 108, all while decreasing the mAP value by only 0.008 for 4 inputs and by only 0.034 for 9 inputs. Considering the obtained FPS values and achieved hardware efficiency, these minimal losses of mAP are easily acceptable.