Human Action Recognition with OpenVINO

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS. To run without installing anything, click the launch binder button.

Binder Github

This notebook demonstrates live human action recognition with OpenVINO. We use the Action Recognition Models from Open Model Zoo, specifically an Encoder and a Decoder. Both models create a sequence to sequence ("seq2seq")1system to identify the human activities for Kinetics-400 dataset. The models use the Video Transformer approach with ResNet34 encoder2. In the notebook we show how to create the following pipeline:

At the end of this notebook, you will see live inference results from your webcam. You can also upload a video file.

NOTE: To use the webcam, you must run this Jupyter notebook on a computer with a webcam. If you run on a server, the webcam will not work. However, you can still do inference on a video in the final step.

1 seq2seq: Deep learning models that take a sequence of items to the input and output. In this case, input: video frames, output: actions sequence. This "seq2seq" is composed of an encoder and a decoder. The encoder captures the "context" of the inputs to be analyzed by the decoder, and finally gets the human action and confidence.:math:hookleftarrow` <#a1>`__

2 Video Transformer and ResNet34. :math:hookleftarrow` <#a2>`__


import collections
import os
import sys
import time
from typing import Tuple, List

import cv2
import numpy as np
from IPython import display
from openvino.runtime import Core
from openvino.runtime.ie_api import CompiledModel

import notebook_utils as utils

The models

Download the models

We use omz_downloader, which is a command-line tool from the openvino-dev package. omz_downloader automatically creates a directory structure and downloads the selected model.

In this case you can use "action-recognition-0001" as a model name, and the system automatically downloads the two models "action-recognition-0001-encoder" and "action-recognition-0001-decoder"

Note: If you want to download another model, such as ``”driver-action-recognition-adas-0002”`` (``”driver-action-recognition-adas-0002-encoder”`` + ``”driver-action-recognition-adas-0002-decoder”``), please change the model name. Using a model outside the list can require different pre- and post-processing.

# Directory where model will be downloaded
base_model_dir = "model"
# Model name as named in Open Model Zoo
model_name = "action-recognition-0001"
# Selected precision (FP32, FP16, FP16-INT8)
precision = "FP16"
model_path_decoder = (
model_path_encoder = (
if not os.path.exists(model_path_decoder) or not os.path.exists(model_path_encoder):
    download_command = f"omz_downloader " \
                       f"--name {model_name} " \
                       f"--precision {precision} " \
                       f"--output_dir {base_model_dir}"
    ! $download_command
################|| Downloading action-recognition-0001-decoder ||################

========== Downloading model/intel/action-recognition-0001/action-recognition-0001-decoder/FP16/action-recognition-0001-decoder.xml

========== Downloading model/intel/action-recognition-0001/action-recognition-0001-decoder/FP16/action-recognition-0001-decoder.bin

################|| Downloading action-recognition-0001-encoder ||################

========== Downloading model/intel/action-recognition-0001/action-recognition-0001-encoder/FP16/action-recognition-0001-encoder.xml

========== Downloading model/intel/action-recognition-0001/action-recognition-0001-encoder/FP16/action-recognition-0001-encoder.bin

Load your labels

We are using Kinetics-400 dataset, and also we are providing you with the text file embedded into this notebook.

Note: If you want to run ``”driver-action-recognition-adas-0002”`` model, change the txt file for ``”driver_actions.txt”``

labels = "data/kinetics.txt"

with open(labels) as f:
    labels = [line.strip() for line in f]

print(labels[0:9], np.shape(labels))
['abseiling', 'air drumming', 'answering questions', 'applauding', 'applying cream', 'archery', 'arm wrestling', 'arranging flowers', 'assembling computer'] (400,)

Load the models

We will load the two models for this particular architecture, Encoder and Decoder. Downloaded models are located in a fixed structure, indicating vendor, model name, and precision.

  1. Initialize inference engine (IECore)

  2. Read the network from .bin and.xml files (weights and architecture)

  3. Compile the model for the “CPU.”

  4. Get input and output names of nodes.

Only a few lines of code are required to run the model. Let’s see it.

Model Initialization function

# Initialize inference engine
ie_core = Core()

def model_init(model_path: str) -> Tuple:
    Read the network and weights from file, load the
    model on the CPU and get input and output names of nodes

    :param: model: model architecture path *.xml
            compiled_model: Compiled model
            input_key: Input node for model
            output_key: Output node for model

    # Read the network and corresponding weights from file
    model = ie_core.read_model(model=model_path)
    # compile the model for the CPU (you can use GPU or MYRIAD as well)
    compiled_model = ie_core.compile_model(model=model, device_name="CPU")
    # Get input and output names of nodes
    input_keys = compiled_model.input(0)
    output_keys = compiled_model.output(0)
    return input_keys, output_keys, compiled_model

Initialization for Encoder and Decoder

# Encoder initialization
input_key_en, output_keys_en, compiled_model_en = model_init(model_path_encoder)
# Decoder initialization
input_key_de, output_keys_de, compiled_model_de = model_init(model_path_decoder)

# Get input size - Encoder
height_en, width_en = list(input_key_en.shape)[2:]
# Get input size - Decoder
frames2decode = list(input_key_de.shape)[0:][1]

Helper functions

We will use the following helper functions for preprocessing and postprocessing frames:

  1. Preprocess the input image before running the Encoder model. (center_crop and adaptative_resize)

  2. Decode top-3 probabilities into label names. (decode_output)

  3. Draw the Region of Interest (ROI) over the video. (rec_frame_display)

  4. Prepare the frame for displaying label names over the video. (display_text_fnc)

def center_crop(frame: np.ndarray) -> np.ndarray:
    Center crop squared the original frame to standardize the input image to the encoder model

    :param frame: input frame
    :returns: center-crop-squared frame
    img_h, img_w, _ = frame.shape
    min_dim = min(img_h, img_w)
    start_x = int((img_w - min_dim) / 2.0)
    start_y = int((img_h - min_dim) / 2.0)
    roi = [start_y, (start_y + min_dim), start_x, (start_x + min_dim)]
    return frame[start_y : (start_y + min_dim), start_x : (start_x + min_dim), ...], roi

def adaptive_resize(frame: np.ndarray, size: int) -> np.ndarray:
     The frame going to be resized to have a height of size or a width of size

    :param frame: input frame
    :param size: input size to encoder model
    :returns: resized frame, np.array type
    h, w, _ = frame.shape
    scale = size / min(h, w)
    w_scaled, h_scaled = int(w * scale), int(h * scale)
    if w_scaled == w and h_scaled == h:
        return frame
    return cv2.resize(frame, (w_scaled, h_scaled))

def decode_output(probs: np.ndarray, labels: np.ndarray, top_k: int = 3) -> np.ndarray:
    Decodes top probabilities into corresponding label names

    :param probs: confidence vector for 400 actions
    :param labels: list of actions
    :param top_k: The k most probable positions in the list of labels
    :returns: decoded_labels: The k most probable actions from the labels list
              decoded_top_probs: confidence for the k most probable actions
    top_ind = np.argsort(-1 * probs)[:top_k]
    out_label = np.array(labels)[top_ind.astype(int)]
    decoded_labels = [out_label[0][0], out_label[0][1], out_label[0][2]]
    top_probs = np.array(probs)[0][top_ind.astype(int)]
    decoded_top_probs = [top_probs[0][0], top_probs[0][1], top_probs[0][2]]
    return decoded_labels, decoded_top_probs

def rec_frame_display(frame: np.ndarray, roi) -> np.ndarray:
    Draw a rec frame over actual frame

    :param frame: input frame
    :param roi: Region of interest, image section processed by the Encoder
    :returns: frame with drawed shape


    cv2.line(frame, (roi[2] + 3, roi[0] + 3), (roi[2] + 3, roi[0] + 100), (0, 200, 0), 2)
    cv2.line(frame, (roi[2] + 3, roi[0] + 3), (roi[2] + 100, roi[0] + 3), (0, 200, 0), 2)
    cv2.line(frame, (roi[3] - 3, roi[1] - 3), (roi[3] - 3, roi[1] - 100), (0, 200, 0), 2)
    cv2.line(frame, (roi[3] - 3, roi[1] - 3), (roi[3] - 100, roi[1] - 3), (0, 200, 0), 2)
    cv2.line(frame, (roi[3] - 3, roi[0] + 3), (roi[3] - 3, roi[0] + 100), (0, 200, 0), 2)
    cv2.line(frame, (roi[3] - 3, roi[0] + 3), (roi[3] - 100, roi[0] + 3), (0, 200, 0), 2)
    cv2.line(frame, (roi[2] + 3, roi[1] - 3), (roi[2] + 3, roi[1] - 100), (0, 200, 0), 2)
    cv2.line(frame, (roi[2] + 3, roi[1] - 3), (roi[2] + 100, roi[1] - 3), (0, 200, 0), 2)
    # Write ROI over actual frame
    org = (roi[2] + 3, roi[1] - 3)
    org2 = (roi[2] + 2, roi[1] - 2)
    FONT_SIZE = 0.5
    FONT_COLOR = (0, 200, 0)
    FONT_COLOR2 = (0, 0, 0)
    cv2.putText(frame, "ROI", org2, FONT_STYLE, FONT_SIZE, FONT_COLOR2)
    cv2.putText(frame, "ROI", org, FONT_STYLE, FONT_SIZE, FONT_COLOR)
    return frame

def display_text_fnc(frame: np.ndarray, display_text: str, index: int):
    Include text on the analized frame

    :param frame: input frame
    :param display_text: text to add on the frame
    :param index: index line dor adding text

    # Configuration for displaying images with text
    FONT_COLOR = (255, 255, 255)
    FONT_COLOR2 = (0, 0, 0)
    FONT_SIZE = 0.7
    # ROI over actual frame
    (processed, roi) = center_crop(frame)
    # Draw a ROI over actual frame
    frame = rec_frame_display(frame, roi)
    # Put text over actual frame
    text_loc = (TEXT_LEFT_MARGIN, TEXT_VERTICAL_INTERVAL * (index + 1))
    text_loc2 = (TEXT_LEFT_MARGIN + 1, TEXT_VERTICAL_INTERVAL * (index + 1) + 1)
    cv2.putText(frame, display_text, text_loc2, FONT_STYLE, FONT_SIZE, FONT_COLOR2)
    cv2.putText(frame, display_text, text_loc, FONT_STYLE, FONT_SIZE, FONT_COLOR)

AI Functions

Following this pipeline,

we will use the next functions to help us to:

  1. Preprocess frame before running the Encoder. (preprocessing)

  2. Encoder Inference per frame. (encoder)

  3. Decoder inference per set of frames. (decoder)

  4. Normalize the Decoder output to get confidence values per action recognition label. (softmax)

def preprocessing(frame: np.ndarray, size: int) -> np.ndarray:
    Preparing frame before Encoder.
    The image should be scaled to its shortest dimension at "size"
    and cropped, centered, and squared so that both width and
    height have lengths "size". Frame must be transposed from
    Height-Width-Channels (HWC) to Channels-Height-Width (CHW).

    :param frame: input frame
    :param size: input size to encoder model
    :returns: resized and cropped frame
    # Adaptative resize
    preprocessed = adaptive_resize(frame, size)
    # Center_crop
    (preprocessed, roi) = center_crop(preprocessed)
    # Transpose frame HWC -> CHW
    preprocessed = preprocessed.transpose((2, 0, 1))[None,]  # HWC -> CHW
    return preprocessed, roi

def encoder(
    preprocessed: np.ndarray,
    compiled_model: CompiledModel
) -> List:
    Encoder Inference per frame. This function calls the network previously
    configured for the encoder model (compiled_model), extracts the data
    from the output node, and appends it in an array to be used by the decoder.

    :param: preprocessed: preprocessing frame
    :param: compiled_model: Encoder model network
    :returns: encoder_output: embedding layer that is appended with each arriving frame
    output_key_en = compiled_model.output(0)

    # Get results on action-recognition-0001-encoder model
    infer_result_encoder = compiled_model([preprocessed])[output_key_en]
    return infer_result_encoder

def decoder(encoder_output: List, compiled_model_de: CompiledModel) -> List:
    Decoder inference per set of frames. This function concatenates the embedding layer
    froms the encoder output, transpose the array to match with the decoder input size.
    Calls the network previously configured for the decoder model (compiled_model_de), extracts
    the logits and normalize those to get confidence values along specified axis.
    Decodes top probabilities into corresponding label names

    :param: encoder_output: embedding layer for 16 frames
    :param: compiled_model_de: Decoder model network
    :returns: decoded_labels: The k most probable actions from the labels list
              decoded_top_probs: confidence for the k most probable actions
    # Concatenate sample_duration frames in just one array
    decoder_input = np.concatenate(encoder_output, axis=0)
    # Organize input shape vector to the Decoder (shape: [1x16x512]]
    decoder_input = decoder_input.transpose((2, 0, 1, 3))
    decoder_input = np.squeeze(decoder_input, axis=3)
    output_key_de = compiled_model_de.output(0)
    # Get results on action-recognition-0001-decoder model
    result_de = compiled_model_de([decoder_input])[output_key_de]
    # Normalize logits to get confidence values along specified axis
    probs = softmax(result_de - np.max(result_de))
    # Decodes top probabilities into corresponding label names
    decoded_labels, decoded_top_probs = decode_output(probs, labels, top_k=3)
    return decoded_labels, decoded_top_probs

def softmax(x: np.ndarray) -> np.ndarray:
    Normalizes logits to get confidence values along specified axis
    x: np.array, axis=None
    exp = np.exp(x)
    return exp / np.sum(exp, axis=None)

Main Processing Function

Run action recognition function will run in different operations, either a webcam or a video file. See the list of procedures below:

  1. Create a video player to play with target fps (utils.VideoPlayer).

  2. Prepare a set of frames to be encoded-decoded.

  3. Run AI functions

  4. Visualize the results.

def run_action_recognition(
    source: str = "0",
    flip: bool = True,
    use_popup: bool = False,
    compiled_model_en: CompiledModel = compiled_model_en,
    compiled_model_de: CompiledModel = compiled_model_de,
    skip_first_frames: int = 0,
    Use the "source" webcam or video file to run the complete pipeline for action-recognition problem
    1. Create a video player to play with target fps
    2. Prepare a set of frames to be encoded-decoded
    3. Preprocess frame before Encoder
    4. Encoder Inference per frame
    5. Decoder inference per set of frames
    6. Visualize the results

    :param: source: webcam "0" or video path
    :param: flip: to be used by VideoPlayer function for flipping capture image
    :param: use_popup: False for showing encoded frames over this notebook, True for creating a popup window.
    :param: skip_first_frames: Number of frames to skip at the beginning of the video.
    :returns: display video over the notebook or in a popup window

    size = height_en  # Endoder input size - From Cell 5_9
    sample_duration = frames2decode  # Decoder input size - From Cell 5_7
    # Select frames per second of your source
    fps = 30
    player = None
        # Create a video player
        player = utils.VideoPlayer(source, flip=flip, fps=fps, skip_first_frames=skip_first_frames)
        # Start capturing
        if use_popup:
            title = "Press ESC to Exit"
            cv2.namedWindow(title, cv2.WINDOW_GUI_NORMAL | cv2.WINDOW_AUTOSIZE)

        processing_times = collections.deque()
        processing_time = 0
        encoder_output = []
        decoded_labels = [0, 0, 0]
        decoded_top_probs = [0, 0, 0]
        counter = 0
        # Create a text template to show inference results over video
        text_inference_template = "Infer Time:{Time:.1f}ms,{fps:.1f}FPS"
        text_template = "{label},{conf:.2f}%"

        while True:
            counter = counter + 1

            # read a frame from the video stream
            frame =
            if frame is None:
                print("Source ended")

            scale = 1280 / max(frame.shape)

            # Adaptative resize for visualization
            if scale < 1:
                frame = cv2.resize(frame, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)

            # Select one frame every two for processing through the encoder.
            # After 16 frames are processed, the decoder will find the action,
            # and the label will be printed over the frames.

            if counter % 2 == 0:
                # Preprocess frame before Encoder
                (preprocessed, _) = preprocessing(frame, size)

                # Measure processing time
                start_time = time.time()

                # Encoder Inference per frame
                encoder_output.append(encoder(preprocessed, compiled_model_en))

                # Decoder inference per set of frames
                # Wait for sample duration to work with decoder model
                if len(encoder_output) == sample_duration:
                    decoded_labels, decoded_top_probs = decoder(encoder_output, compiled_model_de)
                    encoder_output = []

                # Inference has finished ... Let's to display results
                stop_time = time.time()

                # Calculate processing time
                processing_times.append(stop_time - start_time)

                # Use processing times from last 200 frames
                if len(processing_times) > 200:

                # Mean processing time [ms]
                processing_time = np.mean(processing_times) * 1000
                fps = 1000 / processing_time

            # Visualize the results
            for i in range(0, 3):
                display_text = text_template.format(
                    conf=decoded_top_probs[i] * 100,
                display_text_fnc(frame, display_text, i)

            display_text = text_inference_template.format(Time=processing_time, fps=fps)
            display_text_fnc(frame, display_text, 3)

            # Use this workaround you experience flickering
            if use_popup:
                cv2.imshow(title, frame)
                key = cv2.waitKey(1)
                # escape = 27
                if key == 27:
                # Encode numpy array to jpg
                _, encoded_img = cv2.imencode(".jpg", frame, params=[cv2.IMWRITE_JPEG_QUALITY, 90])
                # Create IPython image
                i = display.Image(data=encoded_img)
                # Display the image in this notebook

    # ctrl-c
    except KeyboardInterrupt:
    # Any different error
    except RuntimeError as e:
        if player is not None:
            # stop capturing
        if use_popup:

Run Action Recognition on a Video File

Now let’s see how the model works in a video file. Any format supported by OpenCV will work. Anytime you can press the stop button while the video file is running, it will activate the webcam for the next step.

Note: Sometimes, the video could be cut off if there are corrupted frames. In that case you could convert. If you experience any problems with your video, please use theHandBrakeand select the MPEG format.

video_file = ""
run_action_recognition(source=video_file, flip=False, use_popup=False, skip_first_frames=600)
Source ended

Run Action Recognition using your webcam

Now let’s try to see ourselves in our webcam.

NOTE: To use the webcam, you must run this Jupyter notebook on a computer with a webcam. If you run on a server, the webcam will not work. However, you can still do inference on a video in the final step.

run_action_recognition(source=0, flip=False, use_popup=False, skip_first_frames=0)
Cannot open camera 0