A High-Throughput Memory Compression Architecture for Real-Time 3D Visual Tracking and Quantification on Edge Devices

Abstract

The deployment of real-time three-dimensional (3D) visual perception, continuous spatial sensing, and graphics-intensive multimedia applications on embedded edge devices introduces substantial runtime memory pressure. During execution, these workloads continuously generate small-granularity memory pages, including image buffers, depth maps, rendering caches, intermediate feature tensors, and media buffers. To reduce memory footprint while preserving low processing overhead, this paper proposes a sequence-aware cascaded compression framework for runtime memory pages generated by visual and multimedia workloads. Instead of following the traditional restore-then-recompress path, the proposed framework directly reuses the intermediate sequence representation produced by LZ4. It parses LZ4 sequences into literal and structured stream sequences and applies stream-specific entropy coding using Huffman and Finite State Entropy (FSE) coding. By avoiding full-page restoration and repeated redundancy analysis, the framework preserves lossless reconstruction semantics while reducing compression overhead. Experimental results on real-world runtime memory pages show that the proposed framework reduces compression processing time by 42.10% on average and by up to 52.22%. It also reduces Central Processing Unit (CPU) time by 20.28% on average and peak memory usage by 29.92% on average. Additional validation on a RISC-V edge platform shows that the compression processing time is reduced by up to 64.36%. A 16 MB point-cloud memory sample derived from 3D scan data further shows a 33.16% reduction in compression time. Functional tests on more than 200K mixed-entropy memory pages demonstrate 100% lossless reconstruction and successful detection of corrupted compressed streams.