How to apply mmaction to live camera for action recognition?

[elepherai]

Hi, I want to do action recognition on live camera in real time? Do you have any suggestions? Thanks~

[kruthikakr]

Looking for any inputs on this ?

[charlielito]

I adapted the demo/demo_inference.py to work with live camera. Here is the code

# Copyright (c) OpenMMLab. All rights reserved.
from argparse import ArgumentParser
import cv2
import numpy as np
from collections import deque
from mmaction.apis.inferencers import MMAction2Inferencer

def parse_args():
    parser = ArgumentParser()
    parser.add_argument(
        'inputs', type=str, help='Input video device')
    parser.add_argument(
        '--vid-out-dir',
        type=str,
        default='',
        help='Output directory of videos.')
    parser.add_argument(
        '--rec',
        type=str,
        default=None,
        help='Pretrained action recognition algorithm. It\'s the path to the '
        'config file or the model name defined in metafile.')
    parser.add_argument(
        '--rec-weights',
        type=str,
        default=None,
        help='Path to the custom checkpoint file of the selected recog model. '
        'If it is not specified and "rec" is a model name of metafile, the '
        'weights will be loaded from metafile.')
    parser.add_argument(
        '--label-file', type=str, default=None, help='label file for dataset.')
    parser.add_argument(
        '--device',
        type=str,
        default=None,
        help='Device used for inference. '
        'If not specified, the available device will be automatically used.')
    parser.add_argument(
        '--batch-size', type=int, default=1, help='Inference batch size.')
    parser.add_argument(
        '--show',
        action='store_true',
        help='Display the video in a popup window.')
    parser.add_argument(
        '--print-result',
        action='store_true',
        help='Whether to print the results.')
    parser.add_argument(
        '--pred-out-file',
        type=str,
        default='',
        help='File to save the inference results.')

    call_args = vars(parser.parse_args())

    init_kws = ['rec', 'rec_weights', 'device', 'label_file']
    init_args = {}
    for init_kw in init_kws:
        init_args[init_kw] = call_args.pop(init_kw)

    return init_args, call_args

def main():
    init_args, call_args = parse_args()
    init_args["input_format"] = "array"
    mmaction2 = MMAction2Inferencer(**init_args)

    video_device = call_args.pop('inputs')
    cap = cv2.VideoCapture(video_device)
    sequence_len  = 5
    frames = deque(maxlen=sequence_len)
    while True:
        ret, frame = cap.read()
        if not ret:
            break
        frames.append(frame)
        if len(frames) == sequence_len:
            inputs = np.array(frames)

            call_args["inputs"] = inputs
            results = mmaction2(**call_args)
            preds = results["predictions"][0]["rec_scores"][0]   
            # argmax of list
            pred_index = np.argmax(preds)

            cv2.putText(frame, pred_index, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
            cv2.imshow("frame", frame)
            cv2.waitKey(1)

if __name__ == '__main__':
    main()

Then you just run with

python script.py /dev/video0 <other-args>

[duylddev]

@charlielito Can we custom the code to support user client webcam instead of connect to device directly from python? eg. any user go to website.com and provide webcam permission can see the real time output video

[charlielito]

@charlielito Can we custom the code to support user client webcam instead of connect to device directly from python? eg. any user go to website.com and provide webcam permission can see the real time output video

No, that's far from straightforward since the python code doesn't have access to the webpage video stream. You'll need a custom bridge to connect the python script to the web-browser like a websocket or webrtc

[duylddev]

If it’s webrtc, how we implement in python side? Can opencv or mmaction can read from webrtc directly?

[charlielito]

If it’s webrtc, how we implement in python side? Can opencv or mmaction can read from webrtc directly?

Nope, you would need to do a custom implementation to interface with mmaction. For example use this library: https://github.com/aiortc/aiortc