Comparative Analysis of Open-Source Video Analytics Systems for Task Distribution in Heterogeneous Computing Environments

Comparative Analysis of Open-Source Video Analytics Systems for Task Distribution in Heterogeneous Computing Environments

Balashov S.A. (MPEI, Moscow, Russia)

Abstract

This paper presents a comparative analysis of eleven open-source video analytics systems from the perspective of task distribution in heterogeneous computing environments combining CPUs and GPUs. The growing volume of video data requires real-time processing on heterogeneous devices (CPU, GPU, VPU, FPGA, TPU), making efficient task scheduling challenging. The platforms examined include NVIDIA DeepStream SDK, Intel OpenVINO with GStreamer Video Analytics, GStreamer, Google MediaPipe, Savant, Deep-Framework, Distream, Frigate NVR, Apache Kafka + Flink, Llama, and Scanner. Task distribution approaches were systematized– from static pipeline assignment to dynamic DAG scheduling. A comparison table across seven criteria was constructed, revealing that no system combines industrial-grade performance with adaptive scheduling. The principal gap is the absence of an adaptive scheduler simultaneously accounting for hardware heterogeneity (CPU+GPU), QoS requirements, workload dynamics, and multi-stream balancing. An experimental study on RTX 3060 + i7-14700K with five systems and a standardized protocol revealed that lightweight implementations (GStreamer custom pipeline and optimized C++ prototype) achieve 2.7× and 8.3× higher throughput than DeepStream videoanalytics SDK, identified significant Python wrapper overhead in Savant (44.6 ms latency vs 2.3 ms for pipeline-based multimedia framework GStreamer), and validated CPU inference offloading via OpenVINO (58.2 FPS with 52% CPU utilization). These findings establish the relevance of developing an adaptive scheduler integrating HEFT/CPOP heuristics, reinforcement learning, and real-time monitoring.

Keywords

video analytics; task distribution; heterogeneous computing; DAG scheduling; GPU acceleration; scalability; open-source; pipeline processing; real-time; quality of service.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 87-96

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(6)-6

For citation

Balashov S.A. Comparative Analysis of Open-Source Video Analytics Systems for Task Distribution in Heterogeneous Computing Environments. Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 87-96 DOI: 10.15514/ISPRAS-2026-38(6)-6.

Full text of the paper in pdf Back to the contents of the volume