# algolive

**Repository Path**: H-kernel/algolive

## Basic Information

- **Project Name**: algolive
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-05-23
- **Last Updated**: 2026-05-23

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# AlgoLive

AlgoLive is a C++ RTSP detection relay with a browser management UI.

## Features

- Pull RTSP streams with configurable TCP or UDP transport.
- Prefer NVIDIA FFmpeg GPU codecs for H264/H265 decode and H264 encode when available.
- Run OpenCV DNN detection on decoded frames.
- Draw detection boxes and labels on the video frame.
- Push H264 fragmented MP4 segments to the browser over WebSocket for MSE playback.
- Optionally push encoded output to a configured RTMP endpoint.
- Manage RTSP channels, live playback, model parameters, and codec defaults from the web UI.

## Build

Install system dependencies:

- CMake 3.20+
- C++20 compiler
- OpenCV with DNN
- FFmpeg development libraries: `libavformat`, `libavcodec`, `libavutil`, `libswscale`

Then build:

```bash
./scripts/build.sh
```

## Run

```bash
./build/algolive --config config/example.json
```

Open `http://127.0.0.1:8080/`.

The example stream is disabled by default. Edit it in the UI or in `config/example.json`, then save.

## Web UI

- `频道管理`: add and edit multiple RTSP channels, channel names, RTSP TCP/UDP mode, detection prompts, RTMP push toggle, RTMP URL, and per-channel encode parameters.
- `实时播放`: select enabled channels and play multiple live H264+fMP4 WebSocket streams with MSE.
- `系统配置`: manage model paths, detector thresholds/input size, service listen settings, and default codec parameters used by new channels.

## Detection Model

Set `detector.model_path` to a YOLO-style ONNX model and `detector.labels_path` to a newline-delimited class label file. The current detector filters class labels by each stream's `prompt` field.

Set `detector.model_type` to `yolo_altclip` to enable AltCLIP semantic filtering. In this mode YOLO still generates candidate boxes, then AltCLIP scores each cropped box against text embeddings and rewrites/filter labels by the stream prompt.

AltCLIP configuration:

- `detector.altclip_model_path`: ONNX image encoder.
- `detector.altclip_embeddings_path`: text embedding file.
- `detector.altclip_threshold`: minimum cosine similarity.
- `detector.altclip_input_width` / `detector.altclip_input_height`: image encoder input size.

The text embedding file accepts one label per line:

```text
person<TAB>0.12 0.34 ...
car,0.11,0.22,...
```