Object Detection

gst-ai-object-detection performs object detection on each frame of a video stream, drawing bounding boxes around detected objects with class labels and confidence scores.

Supports YOLOX, YOLOv8, and YOLO-NAS models.

Prerequisites

• Completed QIM SDK Installation and Model Download

Steps

1. Verify Model and Labels

radxa@airbox$

ls -l /etc/models/yolox_quantized.tflite
ls -l /etc/labels/yolox.json

2. View Configuration

radxa@airbox$

cat /etc/configs/config_detection.json

Key fields:

Field	Default	Description
file-path	/etc/media/video.mp4	Input video path
ml-framework	tflite	Inference framework
yolo-model-type	yolox	YOLO type: yolov5 / yolov8 / yolox / yolonas
model	/etc/models/yolox_quantized.tflite	Model file
labels	/etc/labels/yolox.json	Label file
threshold	40	Confidence threshold (1-100)
runtime	dsp	Inference hardware

3. Run

radxa@airbox$

gst-ai-object-detection --config-file=/etc/configs/config_detection.json

Press Ctrl + C to stop.

Expected Output

Terminal output:

Running app with model: /etc/models/yolox_quantized.tflite and labels: /etc/labels/yolox.json
Using DSP Delegate
VERBOSE: Replacing 329 out of 329 node(s) with delegate (TfLiteQnnDelegate) node
Pipeline state changed from PAUSED to PLAYING

The display shows the test video with colored bounding boxes around detected objects, annotated with class names and confidence scores.

Validation

• Using DSP Delegate: Inference running on NPU
• Replacing 329 out of 329 node(s): All 329 operators delegated to DSP
• Pipeline reaches PLAYING state
• Display correctly shows bounding boxes and labels

Adjusting Detection Sensitivity

If too few objects are detected (missed detections), lower the threshold:

radxa@airbox$

# Change threshold to 20 in /etc/configs/config_detection.json

If there are too many false positives, increase the threshold (e.g., 60).

Switching YOLO Models

Modify the model type and path in the config file:

{
  "yolo-model-type": "yolov8",
  "model": "/etc/models/yolov8_det_quantized.tflite",
  "labels": "/etc/labels/yolov8.json"
}

YOLOv8 and YOLO-NAS models require separate export via Qualcomm AI Hub.

Pipeline Flow

filesrc → qtdemux → h264parse → v4l2h264dec → qtimlvconverter
                                                       ↓
                                          qtimltflite (DSP inference)
                                                       ↓
                                          qtimlvdetection (YOLO post-process)
                                                       ↓
                                                qtivcomposer
                                                       ↓
                                                 waylandsink