佈署基於深度學習物件偵測和追蹤應用，通常需要仰賴雲端的大型主機。隨著邊緣攝影機的數量越來越多，基於雲端運算的 AI 系統，有著頻寬消耗大、延遲長、以及功耗高等缺點，因此導致近年來邊緣運算快技術速崛起。本研究提出一個創新的 AIoT 平台，可用於佈署 AI-enabled 攝影機網路。我們將物件追蹤部署至邊緣端，藉此降低系統的時間延遲。本系統藉由彈性擴增運算單元，針對不同邊緣攝影機數量，快速調整閘道器的運算單元，以便達成藉此達成效能需求和功耗的平衡。透過攝影機網路的協同追蹤，物件在昏暗環境或是物體微小的情況，皆能夠成功偵測和追蹤。我們以行人追蹤作為實驗測試資料，得出我們的平台在網路傳輸效率上相較大型主機節省了 64%的時間，而 AI推論時間上則有 10%之效能提升，加上此一 AIoT 平台彈性架構和低功耗的優勢，非常適合應用於 AI 服務需求經常變動的環境。;Deployment of deep learning-based object detection and tracking applications, which typically rely on cloud-based mainframes. With the increasing number of edge cameras, cloud based AI systems have the disadvantages of high bandwidth consumption, high latency, and high power consumption, resulting in the rapid rise of edge computing technology in recent years. This study proposes an innovative AIoT platform that can be used to deploy AI-enabled camera networks. We deploy object tracking task to the edge to reduce the latency of the system. The system is designed to achieve a balance between performance requirements and power consumption by flexibly scaling up the computing units and quickly adjusting the gateway′s computing units for different numbers of edge cameras. With the collaborative tracking of the camera network, objects can be successfully detected even if they are small or in low light environments. Using pedestrian tracking as our experimental test data, we found that our platform saves 64% of time in network transmission efficiency compared to mainframes, and has a 10% performance improvement in AI inference time. Coupling with the flexible architecture and low power consumption of this AIoT platform, makes it ideal for applications in environments where the demand for AI services change frequently.
