Comparison of Analyze-Then-Compress Methods in Edge-Assisted Visual SLAM

In edge-assisted visual Simultaneous Localization and Mapping (SLAM), mobile devices offload computationally intensive tasks to an edge cloud. The device should transfer its input data the cloud in a resource-efficient way, especially SLAM systems with their very large image frame data. Most state-of-the-art use conventional Analyze-Then-Compress (ATC) approach that pre-analyzes captured frames on (incurring substantial processing latencies), transmits only (several times smaller) pre-analysis results, namely feature representation of so-called key (and corresponding tracking results), edge. We examine two novel transmission methods (ATC workflows) for "functional-split" SLAM: Feature-Representation-Every-Frame (FREF) every edge, performs creation (including tracking); Feature-Representation-Only Key Frames (FROKF) without results tracking. evaluate these ATC terms required network throughput (bandwidth), as well latencies processing, resulting end-to-end latency (from capture at becoming available cloud) via testbed measurements various computing hardware platforms. Compared existing approach, our newly introduced FREF method reduces down one quarter; while FROKF can reduce sixth. Additionally, compression throughput; within method, twentieth compared compression, albeit expense significantly increased compression.