【摘要】 目的 将面向服务器端设计的跟踪算法迁移部署到边缘端能显著降低功耗，具有较高的实用价值。当前基于Transformer的跟踪算法具有明显的性能优势，然而部署在边缘端时，却可能产生较高的延迟。为了解决这个问题，提出了一种面向边缘端的线性分解注意力（linearly decomposed attention,LinDA）结构，可有效降低Transformer的计算量和推理延迟。方法 LinDA将多头注意力近似表示成数据依赖部分和数据无关部分的和：对于数据依赖部分，用简单的向量元素间相乘及求和表示，避免了复杂的转置和矩阵乘法；对于数据无关部分，直接利用统计得到的注意力矩阵，然后加上一个可学习偏置向量。这种分解既具有全局注意力，又保持了数据依赖的优点。为了弥补线性分解带来的精度损失，还设计了一种知识蒸馏方案，它在原始的损失函数上增加了两部分蒸馏损失：1）将真实包围框替换成教师模型预测的包围框作为监督目标，称为硬标签知识蒸馏；2）将教师模型预测得分的相对大小作为监督目标，称为关系匹配知识蒸馏。基于LinDA结构进一步实现了一种面向边缘端的目标跟踪算法LinDATrack，并将其部署在国产边缘计算主机HS240上。结果 在多个公开数据集上进行了评测。实验结果表明，该算法在该计算主机上可达到61.6帧/s的跟踪速度，功耗约79.5 W，功耗仅占服务器端的6.2%，同时其在LaSOT和LaSOT＿ext上的成功率（success rate, SUC）相对于服务器端基线算法SwinTrack-T最多仅下降约1.8%。结论 LinDATrack具有良好的速度和精度平衡，在边缘端具有较大的优势。更多还原

【Abstract】 Objective The transfer and deployment of tracking algorithms designed for server ends to edge ends has high practical value. This transformation leads to a remarkable decrease in energy consumption, particularly in situations where resources are limited. In recent years, tracking algorithms that incorporate the Transformer architecture have achieved considerable progress because of their superior performance. Nonetheless, the adaptation of these algorithms for edge computing often encounters difficulties, primarily because of the increased latency. This latency is attributed to the complex nature of the Transformer’s attention mechanism, which requires extensive computational resources. This issue is addressed by introducing an innovative solution called the linearly decomposed attention(LinDA) module, which is designed expressly for edge computing. By drastically lowering the computational demands and reducing the inference time of the Transformer, the LinDA module facilitates more effective and efficient tracking at the edge end.Method LinDA innovatively approximates the multihead attention mechanism as two components: a data-dependent component and a dataindependent component. For the data-dependent aspect, LinDA adopts a computationally economic approach. Rather than relying on traditional, resource-intensive methods of transposition and matrix multiplication, LinDA employs direct elementwise multiplication and the addition of vectors. This method markedly reduces computational complexity, rendering it exceptionally well suited for edge computing environments where resources are scarce. Regarding the data-independent facet, LinDA integrates a statistically derived attention matrix that encapsulates global contextual insights. This matrix is further refined with a learnable bias vector, enhancing the model’s adaptability and versatility. This decomposition strategy empowers LinDA to achieve good precision and considerable efficiency on devices constrained by limited resources. An advanced knowledge distillation strategy, which plays a crucial role in bolstering the student model’s capabilities, is introduced to mitigate potential compromises in accuracy because of the linear decomposition approach. This strategy encompasses two specialized distillation losses integrated into the baseline loss function, each meticulously designed to capture and convey critical insights from the teacher model to the student model. First, the hard label knowledge distillation technique involves replacing the ground-truth bounding box with the bounding box predicted by the teacher model, which serves as the supervision target for the student model. This method allows the student model to learn directly from the teacher’s discernment, thereby enhancing its predictive precision. Consequently, the student model captures the teacher’s knowledge of the problem, which enables it to yield more accurate predictions. Second, the relation matching knowledge distillation strategy harnesses the relationship between the teacher model’s predictions as the supervisory target. This innovative approach captures the complex relationships among different predictions, such as the relative significance of distinct objects or their spatial interrelations. When this relational knowledge is embedded into the student model during training, the model’s performance can be further improved, rendering it more robust and powerful. In summary, this elaborate knowledge distillation framework successfully imparts the teacher model’s insights into the student model, effectively overcoming the potential precision degradation associated with linear decomposition. This scheme ensures that the student model inherits the teacher’s expertise, thereby enabling it to deliver more precise predictions and attain superior performance. This study further implements an edge-end-oriented object tracking algorithm called LinDATrack, which is based on LinDA and distillation. The algorithm is deployed on the domestic edge computing host HS240.Result Comprehensive experiments of the tracker are conducted across various public datasets to test its performance and capability metrics. The experimental results validate the system’s outstanding tracking speed and good precision. With this computing host, LinDATrack achieves an impressive tracking speed of approximately 62 frames per second, facilitating efficient tracking in real-time settings. Furthermore, the system operates with a power consumption of approximately 79. 5 watts, which represents only 6. 2% of the energy used by the server-end configurations. This dramatic reduction in energy usage underscores the system’s exceptional energy efficiency, positioning it as an ideal choice for deployment in settings with limited resources. In an era where energy conservation and sustainability are increasingly important, this system presents a compelling alternative to more energy-intensive options, contributing significantly to a more sustainable computing landscape. In addition to its remarkable tracking speed and low power consumption, the system also exhibits consistently high tracking accuracy, distinguishing it within the realm of object tracking. Compared with the server-end baseline algorithm SwinTrack-T, the system’s tracking accuracy, as determined by the success rate metric, shows only a slight decrease of approximately 1. 8%. This minor decrease in accuracy reflects the system’s capacity to balance performance with efficiency. The scheme can maintain precise tracking functionality while reducing resource usage, rendering it a versatile solution for a broad spectrum of tracking applications.Conclusion LinDATrack is distinguished by its exceptional balance of speed and accuracy, positioning it as a premier option for object tracking applications. Its performance is marked by efficiency, facilitating realtime tracking that users can rely on. Additionally, LinDATrack demonstrates considerable strengths when deployed at the edge, making it exceptionally well suited for environments with limited resources. This combination of speed, accuracy, and edge-oriented advantages firmly establishes LinDATrack as a leading solution for edge-end tracking tasks.更多还原