物体/目标检测器

检测器决定目标推理运行在哪类 AI 硬件或运行时上。

选择建议

后端	适合场景	备注
Coral EdgeTPU	低功耗稳定检测	适合 SSD EdgeTPU 模型
OpenVINO	Intel 平台	CPU / iGPU 视模型支持而定
ONNX / YOLO	自定义模型	配合 Local Training 使用
TensorRT	NVIDIA GPU	需要匹配 CUDA / TensorRT 环境

Coral PCIe

detectors:
  coral:
    type: edgetpu
    device: pci

OpenVINO

detectors:
  ov:
    type: openvino
    device: CPU

ONNX

detectors:
  onnx:
    type: onnx

Object Detectors

The object detector runs neural-network inference. Hardware, runtime, model format, tensor layout, pixel format, data type, input dimensions, and label map must agree. A detector process starting successfully does not prove the model configuration is correct.

Compatibility matrix

Item	Must match
runtime / type	OpenVINO, EdgeTPU, ONNX, TensorRT, RKNN, Hailo, etc.
model path	A model format supported by that runtime.
width / height	The model input shape.
input_tensor	For example NHWC or NCHW.
input_pixel_format	For example RGB or BGR.
input_dtype	Float/int type expected by the model.
labelmap	Class indices and order used during training.

OpenVINO example

detectors:
  ov:
    type: openvino
    device: CPU

model:
  width: 300
  height: 300
  input_tensor: nhwc
  input_pixel_format: bgr
  path: /openvino-model/ssdlite_mobilenet_v2.xml
  labelmap_path: /openvino-model/coco_91cl_bkgr.txt

Performance

Watch inference speed, detector utilization, and detect FPS. If inference cannot keep up across cameras, reduce detection workload, choose a faster model, or add/change acceleration.

Troubleshooting order

Confirm the device is visible inside the container.
Confirm the runtime loads the model without shape/dtype/layer errors.
Verify label-map ordering.
Test one low-FPS camera before scaling out.

CPU detector

CPU inference is useful for testing and small workloads. Multi-camera production systems generally benefit from dedicated inference hardware or GPU/NPU acceleration so inference does not compete with video decoding.

Detectores de objetos

El detector ejecuta la inferencia de la red neuronal. Hardware, runtime, formato de modelo, layout del tensor, formato de píxel, tipo de datos, dimensiones y label map deben coincidir.

Compatibilidad

Elemento	Debe coincidir
runtime / type	OpenVINO, EdgeTPU, ONNX, TensorRT, RKNN, Hailo, etc.
model path	Formato admitido por el runtime.
width / height	Dimensiones de entrada.
input_tensor	NHWC o NCHW.
input_pixel_format	RGB o BGR.
input_dtype	Tipo esperado.
labelmap	Índices/orden de clases.

Ejemplo OpenVINO

detectors:
  ov:
    type: openvino
    device: CPU

model:
  width: 300
  height: 300
  input_tensor: nhwc
  input_pixel_format: bgr
  path: /openvino-model/ssdlite_mobilenet_v2.xml
  labelmap_path: /openvino-model/coco_91cl_bkgr.txt

Rendimiento

Observe inference speed, utilización y detect FPS. Si no se mantiene el ritmo, reduzca carga o use un acelerador/modelo más eficiente.

Diagnóstico

Compruebe el dispositivo dentro del contenedor.
Compruebe carga del modelo.
Verifique label map.
Pruebe una cámara a bajo FPS antes de escalar.

CPU

CPU sirve para pruebas y cargas pequeñas; varias cámaras suelen beneficiarse de aceleración dedicada.