一种目标识别推理平台设计

Design of a Target Recognition and Reasoning Platform

本文从目标识别的应用需求和发展趋势出发,结合硬、软件技术设计了一种推理端的实现平台。硬件采用复旦微电子FPGA-JFM7K325T作为多视频接口适配与进行预处理,有效提高了平台兼容性与时效性,同时采用海思Hi3559A作为目标识别推理算法的核心处理器,在满足既定算法算力的条件下,有效降低了平台体积与功耗,达到小型轻量化目标。软件方面,为提高目标识别置信度,采用预处理技术提高视频质量,并采用基于YOLOv3的改进网络结构,进行目标识别算法的推演;同时结合多目标动态特征,实现目标跟踪算法,扩大平台实际应用场景。本平台解决了推理系统中小型化瓶颈问题,同步提高了小目标识别置信度,适用于航空机载、海军舰载以及陆军战车侦察等环境下的自主侦察、识别应用。

Starting from the application requirements and development trends of the target recognition,this paper proposed an inference platform in combination with hardware and software technology.In terms of hardware,Fudan Microelectronics FPGA JFM7K325T was used as the multi-video interface adaptation and preprocessing,which effectively improved platform compatibility and timeliness.This platform used HiSilicon Hi3559A as the core processor of the target recognition inference algorithm,with the condition of meeting the established algorithm power,effectively reduced the size and power consumption,and reached the goal of small and lightweight.In terms of software,in order to improve the confidence of target recognition,preprocessing technology was used to improve video quality,and then the improved network structure based on YOLOv3 was used to perform target recognition algorithm deduction.At the same time,combined with multi-target dynamic characteristics,target tracking algorithm was realized,and the practical application of the platform was expanded.In summary,this platform solved the problem of miniaturization in the reasoning system,and improved the confidence of small target recognition.It was suitable for autonomous reconnaissance and recognition applications in aviation airborne,navy shipborne and army tank reconnaissance environments.