Real-Time Video Streaming and Per-Frame Analysis with Qwen2.5-VL for Maritime Surveillance

Real-time video analytics can greatly enhance maritime domain awareness by providing timely insights for applications such as environmental surveillance and vessel monitoring. This paper presents a system for real-time video streaming and per-frame analysis using the Qwen2.5-VL visual-language model (VLM) via API. We describe the system architecture comprising a Python-based video capture and frame queuing pipeline coupled with the Qwen2.5-VL model for image understanding. Both cloud-based and cloud-edge hybrid deployment scenarios are considered: in the cloud, high-performance GPU inference on the VLM achieves low per-frame latency, whereas in edge deployments with limited compute, frames are selectively offloaded to the cloud. We develop theoretical models for key performance indicators including perframe analysis time, end-to-end latency, and frame drop rate, and we validate these with empirical measurements from system logs. In an experimental evaluation, the system processes video frames in near real-time by sampling at approximately 1 frame per second (fps) from a 23fps stream. This achieves manageable latency at the cost of reduced frame fidelity. Results demonstrate the trade-offs between analyzing all frames versus skipping frames for performance, and we discuss strategies to balance timely detection of maritime events (such as boats or pollution) with system constraints.