针对传统孪生目标跟踪算法体量大、难以在嵌入式设备部署以及其在目标尺度变化大、有相似物干扰等条件下效果不佳的问题，提出一种新的轻量化快速跟踪(Ghost fast Tracking with TiFPN and Retriever，GTtracker)算法。引入Ghost机制，对Resnet网络进行重新设计，构建一种轻量化G-Resnet网络对跟踪目标进行快速特征提取。设计轻量自适应加权融合(Tiny adaptive weigh-ted fusion algorithm Feature Pyramid Network，TiFPN)算法，进一步加强特征信息的融合，解决相似物干扰问题。设计一种轻量化区域回归网络(Ghost Decoupled Net，GDnet)，以实现目标分类、交并比(Intersection-over-Union，IoU)计算以及边界框回归，并在跟踪阶段应用一种新的目标寻回器提升算法跟踪的成功率。在OTB100数据集和VOT2020数据集上进行算法验证，并移植算法到嵌入式设备Jetson Xavier NX上进行性能测试。实验结果均表明算法的有效性和优越性，相比经典孪生目标跟踪(SiamCAR)算法，新方法在精度和期望平均重叠率(Expected Average Overlap，EAO)指标均相似的前提下，能够实现更快的运行速度，可在Jetson Xavier NX上实时运行，达到30帧/s，且能有效解决相似物干扰、尺度变化大等问题。
Lightweight and Fast Target Tracking Algorithm Based on Ghost-TiFPN
In response to the problems that traditional siamese target tracking algorithm is bulky and difficult to deploy in embedded devices, and has poor effect under the conditions of large changes in target scale and similar object interference, a lightweight and fast target tracking algorithm GTtracker is proposed. The Resnet network is redesigned to build a lightweight G-Resnet network by introducing the Ghost mechanism for fast feature extraction of tracked targets. And then the fusion of feature information is further enhanced by designing a lightweight adaptive weighted fusion algorithm TiFPN to solve the problem of similar object interference. After that, a lightweight area regression network GDNet is introduced for target classification, IoU calculation, and bounding box regression, which applies a new target finder in the tracking stage to enhance the success rate of algorithm tracking. Finally, the algorithm is validated on OTB100 dataset and VOT2020 dataset, and ported to Jetson Xavier NX embedded device for performance testing. Experimental results show the effectiveness and superiority of the proposed algorithm, In comparison with classical siamese target tracking algorithm ( SiamCAR ) , the proposed algorithm can achieve faster operation speed and real-time operation on Jetson Xavier NX, reaching 30 frames/s, under the conditions of the same accuracy and EAO metrics, which can effectively solve the problems of similar object interference and large scale variation.
万方数据
维普
