TriEdge : an integrated architecture for balancing the trilemma in Edge-AI (本文)
Year: 2026
Type: dissertation
Abstract: This dissertation addresses the trilemma among accuracy, real-time performance, and computational resources in Edge-AI, and defines TriEdge, an integrated Edge-AI architecture designed to mitigate this trilemma by moving the balance point among the three axes according to operational conditions.The trilemma cannot be eliminated, but it can be mitigated by design.Trading quality for time and resources is itself classical-imprecise computation, anytime algorithms, voltage scaling-but that lineage optimizes one axis at a time and fixes the operating point at design time for a fixed task with a known workload; TriEdge instead treats the three axes as one trilemma and is designed so that the balance point can be moved rather than fixed offline.Sparse video, in which detection targets appear infrequently, is the verification condition, and urban sensing, including garbage collection, is the application; the target task is garbage bag detection, tracking, and counting from garbage-truck rear cameras, with design principles conditioned on temporal sparsity, intended to generalize to urban video analytics.TriEdge consists of perception, performance tuning, cooperative distribution, and operation functions, and separates data flow from control flow.The functions are abstract definitions, and DeepCounter and JumpQ are replaceable implementation examples.As an implementation of the perception function, DeepCounter realizes on-device detection and counting via a DTC pipeline (Detection/Tracking/Counting), securing the accuracy axis.As an implementation of the performance tuning function, JumpQ is a probabilistic scheduler that lowers detection frequency during consecutive negative intervals and retrospectively processes buffered frames after positive detections, improving the real-time and computational-resource axes while preserving accuracy.The performance tuning function can also be realized by task offloading or model switching.Using operational logs from Fujisawa City, Kanagawa Prefecture (640×480, about 8.8 fps, positive-frame ratio about 12%, 8 vehicles), DeepCounter attains AP 78.9 at 15.2 FPS on Jetson TX2 and counting F1 0.74, the highest among the four detector-tracker configurations compared-higher detection AP does not carry over to counting accuracy.In the integrated setup, JumpQ cuts the number of inferences by about half, raising throughput about 1.92× and lowering average GPU utilization from 62% to 36% while counting accuracy is largely preserved, and the effect is most pronounced under high sparsity.Empirical validation covers two of the four functions, namely perception and performance tuning; the runtime, load-linked mode transitions and cooperative distribution are contributed at the design level, and foundation-model integration is positioned as a design guideline and theoretical estimate with compression, split execution, caching, and control, without measurements.i Evaluation is limited to a specific region and period, and JumpQ depends on prior detection logs, leaving a preparation cost at deployment.Overall, the dissertation systematizes the design and evaluation of sparse-video Edge-AI and provides practical guidance for balancing the trilemma in edge deployment.
Author: ミカミ, カズヒロ
Institution: Keio University
Language: Japanese
Topic: IoT and Edge/Fog Computing
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222386447
