基于NPU的实时深度学习跟踪算法实现

Implementation of real-time deep learning tracking algorithm based on NPU

深度学习在检测领域高速发展,但受限于训练数据和计算效率,在基于嵌入式平台的边缘计算领域,尤其是实时跟踪应用中深度学习的智能化算法应用并不广泛。针对这一现象,同时为满足现阶段国产化、智能化的技术需求,提出了一种改进的孪生网络深度学习跟踪算法。在特征网络加入微调网络,解决了网络模型无法在线更新的问题,提升了跟踪的准确性;在IoUNet损失函数中加入中心距离惩罚项,解决了IoUNet当IoU相同时位置跳跃,存在收敛盲区和收敛速度慢的问题;将训练后的网络通过通道剪枝,缩减网络模型尺寸,提升了模型加载和运行的速度。在华为Atlas200NPU平台上实现了实时运行,算法准确率高达0.90(IoU>0.7),帧率达到66 Hz。

The deep learning has developed rapidly in the field of detection,but limited by the training data and computing efficiency,the intelligent algorithms of deep learning are not widely used in the edge computing field based on embedded platform,especially in real-time tracking applications.Aiming at this phenomenon,in order to meet the needs of domestic and intelligent technology at the present stage,an improved twin network deep learning tracking algorithm was implemented.The fine-tuning network was added to the feature network to solve the problem that the network model could not be updated online and improve the accuracy of tracking.The center distance penalty term was added into IoUNet loss function to solve the problems of position jumping,existence of convergence blind area and the slow convergence when IoU was the same.The trained network was pruned through channels to reduce the size of network model and improved the loading and running speed of the model.Finally,the model was implemented in real-time on Huawei Atlas200NPU platform.The proposed algorithm accuracy is up to 0.90(IoU>0.7),and the frame rate reaches 66 Hz.