Energy-efficient and Accurate Object Detection Design on an FPGA Platform

With the innovation of hardware equipment, the development of artificial intelligence has broken through the limitations of the past. Neural networks have been continuously deepened to improve the accuracy of detection, so that the parameters have increased with a direct proportional rate. In this way, however, high energy consumption has been induced which obstacles the deployment of AI algorithms on portable devices. Therefore, the design of neural network must consider not only detection accuracy but also energy efficiency. In this paper, we analyzed energy consumption, detection accuracy and execution speed of our neural network model as well as the state-of-the-art models based on an FPGA platform called ZCU-102. We adopt the performance index from Low Power Computer Vision (LPCV) challenge which considers power dissipation, mean Average Precision (mAP) and Frames Per Second (FPS) at the same time to evaluate these models in an overall point of view. Agilev4 can achieve 59.9% of mAP@50 on MS COCO test-dev2017 datasets. If the input frame resolution is turned into $416\times 416$, the processing frame rate can reach 20.7 FPS on ZCU-102. Compared with the state-of-the-art models, the LPCV score of Agilev4-416 is 1475. S which is 1.56 times of that of YOLOv4-416.