Object detector for YOLO models

Contents

Object detector for YOLO models

Allows to use YOLO object detection models for image analysis

Installing/updating

Object detector for YOLO models is not included into EVA ICS distribution. To install/update it, either edit “eva/config/python-venv” registry key, specify the desired version in “extra” section (e.g. eva4-svc-yolo-detector>=0.0.1) and rebuild the Python virtual environment (/opt/eva4/sbin/venvmgr build). Or execute:

/opt/eva4/sbin/venvmgr add eva4-svc-yolo-detector
# or
/opt/eva4/sbin/venvmgr add eva4-svc-yolo-detector==N # where N = version number

The latest eva-shell version number can be obtained from https://pypi.org/project/eva4-svc-yolo-detector/

Setup

Use the template EVA_DIR/share/svc-tpl/svc-tpl-yolo-detector.yml:

# Detector service for Ultralytics YOLO models
#
# Requires https://pypi.org/project/eva4-svc-yolo-detector/ Python module installed
command: venv/bin/eva4-svc-yolo-detector
bus:
  path: var/bus.ipc
config:
  # Path to the YOLO model file. In case of model formats other than PyTorch,
  # ensure the YOLO runtime is installed in the venv
  model: /opt/eva4/cv/yolo11m.pt
  # Classes to detect
  classes:
  # The class color used for the debug stream and 3rd party apps
  - color: orange
    # Confidence threshold for the class
    confidence: 0.1
    # Class id, according to the model, can be an integer or a string
    id: person
  # Key points, e.g. for YOLO-pose models. If the color is not specified, the
  # key point is ignored in the debug stream. The model MUST provide keypoints otherwise
  # the detector will fail to work
  #keypoints:
    #colors:
    #- null
    #- green
    #- green
    #- null
    #- null
    #- blue
    #- blue
    #- '#336699'
    #- '#336699'
    #- '#99CCFF'
    #- '#99CCFF'
    #- red
    #- red
    #- '#996633'
    #- '#996633'
    #- '#FFCC99'
    #- '#FFCC99'
    #enable: true
  # Raw stream source OID, must provide RGB8 frames
  oid_src: sensor:streams/s0/raw
  # Debug stream OID, filled with RGB8 frames with detection boxes and keypoints
  #oid_cv_debug: sensor:streams/s0/cv/debug
  # Destination OID for detection results, filled with detection results (object)
  #oid_dst: sensor:streams/s0/cv/json
  # lmacro to process detections, executed with an argument (_1) containing the detections
  # and "detector_stats" variable containing detector statistics
  #process: lmacro:streams/s0/cv/process
user: eva # make sure the user has access to the model file
# the service may consume relatively large amounts of memory, set the warning
# threshold accordingly
mem_warn: 1500000000

Create the service using eva-shell:

eva svc create eva.detector.1 /opt/eva4/share/svc-tpl/svc-tpl-yolo-detector.yml

or using the bus CLI client:

cd /opt/eva4
cat DEPLOY.yml | ./bin/yml2mp | \
    ./sbin/bus ./var/bus.ipc rpc call eva.core svc.deploy -

(see eva.core::svc.deploy for more info)

EAPI methods

See EAPI commons for the common information about the bus, types, errors and RPC calls.

detector.stats

Description	Get detector stats
Parameters	none
Returns	Detector stats

Return payload example:

{
  "fps": 30,
  "frames_dropped": 3,
  "frames_processed": 149,
  "status": "ready",
  "t_avg_ms": 25,
  "t_max_ms": 141,
  "t_min_ms": 18
}