Memory-Efficient Spatiotemporal Feature Representation for Scalable 3D Video Analytics in Information Systems

• Rafiul Hasan

  Patuakhali Science and Technology University, Dumki Road No. 21, Department of Computer Science, Patuakhali, Barisal Division, Bangladesh

  Author

• Siam Chowdhury

  Bangladesh University of Business and Technology, Mirpur Road No. 32, Department of Computer Systems, Dhaka, Dhaka Division, Bangladesh

  Author

• Ariful Islam

  Dhaka, Bangladesh

  Author

Three-dimensional video streams acquired from multi-camera rigs, RGB-D sensors, and volumetric capture systems are increasingly integrated into enterprise workflows, ranging from retail telemetry and industrial inspection to human-computer interaction logs. In information systems settings, the dominant constraints are often not model capacity but memory footprint, ingestion bandwidth, and the need to support continuous queries over long temporal horizons. Conventional spatiotemporal representations derived from dense 3D convolutional backbones or transformer tokenizations can be prohibitively expensive when features must be retained for indexing, auditing, and retrospective analytics, especially under multi-tenant resource governance. This paper develops a memory-efficient spatiotemporal feature representation designed for scalable 3D video analytics in information systems. The approach combines factorized spatiotemporal operators, streaming-safe temporal aggregation, and compressed latent storage that preserves query-relevant similarity structure while controlling memory growth. A formal objective couples representation fidelity with a system-level memory budget, enabling explicit trade-offs among precision, temporal resolution, and retrieval accuracy. The proposed representation supports both real-time inference and offline query execution by decoupling compute-heavy encoding from storage-optimized feature materialization. Analytical modeling characterizes memory, bandwidth, and latency under realistic ingestion and query workloads, while a methodology for end-to-end evaluation is presented across detection, tracking, and similarity search tasks. The result is a principled design space for feature materialization that aligns spatiotemporal learning with database-style constraints on storage, indexing, and throughput.

• PDF

Copyright (c) 2026 International Journal of Science, Engineering and Applied Technologies

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.