面向资源受限无人系统的深度神经网络轻量化软件设计与应用

Design and Application of a Lightweight Deep Neural Network Software for Resource Constrained Unmanned Systems

地外探测无人系统具有存储、算力和能量等资源受限的特点.以深度学习为基础的感知、定位和决策算法可有效提升无人系统的智能化水平,而这类算法通常需要高算力,难以直接应用于地外探测无人系统.首先针对剪枝和量化的深度神经网络模型轻量化方法,在公开数据集上对多种算法进行定量分析.其次,提出基于剪枝、量化的轻量化计算方案,实现了基于模块化配置的轻量化计算软件StarLight,对深度神经网络进行快速轻量化和性能评估,解决了模型难以直接应用到计算资源受限系统的问题.最后,基于StarLight,对应用于火星车实验系统中的多种任务模型进行轻量化,在计算功耗≤15 W、计算处理主频≤1.2 GHz和计算存储容量≤1TB的受限资源条件下,实现了深度神经网络模型部署.实验表明,该软件能够满足计算资源受限系统的深度神经网络模型轻量化需求,为进一步提升地外探测无人系统的智能化水平奠定了基础.

The unmanned system in interplanetary exploration has the characteristics of limited storage,computing power,energy and so on.The perception,localization and decision-making algorithms based on deep neural network can effectively improve the intelligence level,but these algorithms generally require huge computing power,which is difficult to be directly applied to unmanned systems.Therefore,this paper reviews the existing lightweight methods including pruning and quantization,and makes a quantitative analysis on public dataset.Furthermore,this paper proposes pruning and quantization solutions,establishes a lightweight computing software StarLight,realizes rapid lightweight and evaluation of deep neural network,and solves the problem that the deep model is difficult to be directly applied to resource constrained systems.Finally,based on StarLight,various models used in the Mars rover are compressed,and deployed in the embedded platform;under the premise of ensuring performance,the power≤15 W,CPU frequency≤1.2 GHz and storage≤1 TB.Experiments show that the software can meet the lightweight requirements of resource constrained systems,and builds a foundation for further improving the intelligent level of unmanned systems for interplanetary exploration.