Frequently asked questions
Product
How is VSL different from an NVR?
An NVR primarily records footage and retrieves it by time. VSL adds tracked objects, trajectories, rules, VLM summaries, and vectors so users can retrieve events by describing what happened.
Which objects can VSL recognize?
The default COCO models recognize people, cars, trucks, buses, bicycles, animals, and common objects. Capability depends on model training data; a prompt alone cannot make an untrained detector reliably recognize a new category.
Why use a vision-language model?
YOLO and D-FINE are good at identifying visible objects. A VLM is better at describing relationships, activities, anomalies, and risk. VSL uses local detection to select events before requesting deeper semantic analysis.
Privacy and data
Is the full camera stream uploaded?
No. Continuous detection and rule evaluation are local. Only matched events are sent to a VLM when that feature is enabled. Disabling the VLM still leaves detection, tracking, recording, and label-based search operational.
What is stored in Qdrant?
Qdrant stores event vectors and retrieval identifiers. SQLite stores structured event records and job state. Clips, covers, and trajectories remain in the local data directory.
Performance
Can VSL run on CPU?
Yes, for evaluation or low inference rates. NVIDIA CUDA is recommended for sustained real-time analysis.
Does the current version support four cameras?
The current release is a single-stream foundation. Four streams require a multi-camera manager with independent tracking, rules, and recording state per source, ideally sharing one batched inference engine.
Can it run directly on a Jetson Nano?
No. The current Docker image targets x86-64 ONNX Runtime. Jetson requires an ARM64 image, JetPack/TensorRT, and accelerated video decoding. The original 4GB Jetson Nano is not recommended for multi-stream real-time inference.
License
VSL is licensed under GNU AGPL-3.0. You may study, modify, and self-host it subject to the license. See THIRD_PARTY_NOTICES.md for third-party model and runtime terms.