How to use deepstream to detect and capture faces

bvgohmslf · December 7, 2020, 12:01pm

Hi,

How can I detect faces using deepstream while using rtsp from a raspberry pi on jetson nano

DaneLLL · December 8, 2020, 3:46am

Hi,
By default it demonstreates ResNet/SSd/Yolo in DeepStream SDK. You would need to convert the face-recognition model to be TensorRT-runnable and modify the config file to replace the default models.

bvgohmslf · December 8, 2020, 10:46am

How can we do this with dlib since we we have more experience with this and have an implementation with this currently running but having serious lags.

Can you help with a sample detection model I can make use of and how to to implement with tensorRT.

Lastly, we want to connect this to connect this to our python script which records and uploads images and videos to the cloud.

Thanks

AastaLLL · December 9, 2020, 3:22am

Hi,

First, please check if you have enabled GPU support in the dlib library.

If you want some optimization on the camera part from deepstream.
You can check this sample to get the data buffer and feed it into dlib:

github.com

NVIDIA-AI-IOT/deepstream_python_apps/blob/master/apps/deepstream-imagedata-multistream/deepstream_imagedata-multistream.py#L70


      
          TILED_OUTPUT_HEIGHT = 1080
          GST_CAPS_FEATURES_NVMM = "memory:NVMM"
          pgie_classes_str = ["Vehicle", "TwoWheeler", "Person", "RoadSign"]
          
          MIN_CONFIDENCE = 0.3
          MAX_CONFIDENCE = 0.4
          
          
          # tiler_sink_pad_buffer_probe  will extract metadata received on tiler src pad
          # and update params for drawing rectangle, object information etc.
          def tiler_sink_pad_buffer_probe(pad, info, u_data):
              frame_number = 0
              num_rects = 0
              gst_buffer = info.get_buffer()
              if not gst_buffer:
                  print("Unable to get GstBuffer ")
                  return
          
              # Retrieve batch metadata from the gst_buffer
              # Note that pyds.gst_buffer_get_nvds_batch_meta() expects the
              # C address of gst_buffer as input, which is obtained with hash(gst_buffer)

A better solution is to use Deepstream for the whole pipeline.
Suppose you are using the dlib model from this repository:

Then you will need to convert it into TensorRT support format: caffe, ONNX, uff, or directly create the weight with TensorRT API.
You can find the sample of each format in the below folder:

$ /usr/src/tensorrt/samples/

Thanks