Design of Agilev4 and Its Real-time Implementation on an Economic Mobile Platform

Abstract

AI visual algorithms have obtained obvious improvements in detection performance. However, the resulting extremely high computational complexity hinders the progress of AI hardware design for real-time and portable applications. In this paper, we have developed a dedicated AI detection neural network along with the necessary AI model speedup techniques. For the development goal, we propose our own backbone neural network, which we named it Agilev4. Besides, we introduce a so-called GoP (Group of Pictures) technology to accelerate the computation by removing temporal redundancy. The test with a dedicated hyper-market scenario data illustrates that the presented system can track each object well. Finally, the frame rate can be accelerated to 35.4 frames per second when we deployed Agilev4 along with the GoP mode acceleration on Jetson Nano. Compared with the original existing neural network, the speedup ratio is more than 1200%.