一种基于深度学习的目标跟踪加速器

A deep learning object tracking accelerator

针对当前神经网络加速器难以高效实现目标跟踪边框后处理的问题,提出一种高效的目标跟踪专用加速器.引入神经网络架构,用于提取输入视图特征并生成边框置信度与偏移量集合.随后针对目标跟踪的边框处理设计了专用于边框的回归、惩罚以及提取操作的加速模块,通过同步神经网络加速器与专用加速模块间的数据,以流水结构并行执行特征提取与边框操作,实现基于深度学习目标跟踪的端到端处理.该加速器在40 nm工艺下消耗面积3.64mm^(2),获得了5.71 Tops/W能效比.实验结果表明:与现有加速方案相比,该目标跟踪加速器获得了1.53倍加速,可实现实时的视频处理(31 fps).其中仅针对跟踪过程的后处理任务,专用加速模块相对RISC处理器可实现3.2倍的加速比.

Since the current nerual network accelerator couldn t efficiently accelerate the post-processing of object tracking»a dedicated object trackeris proposed.A neural network architecture is introduced to extract the features of the input feature map.At the meanwhile,it generates thebounding box confidence and position offset sets.Adedicated acceleration module is designed for the anchor regression,penalty calculation and extraction.By synchronizing the data between the neural network accelerator and the dedicated module,a new pipelined structure is proposed to execute the feature extraction and anchor regression in parallel.Therefore,the end-to-end processing of the object tracking is efficiently achieved.The accelerator consumes an area of 3.64 mm^(2)under the SMIC 40nm process,and achieves 5.71 Tops/W energy efficiency.Experimental results show that,compared with the current accleration solutions,the object tracking accelerator achieves 1.53 times acceleration,and it could realize real-time video processing(31 fps).For the post-processing task of the tracking,the processing speeds of the proposed dedicated module is improved by 3.2 times than the RISC processor.