Hand Gesture Detection - Rev #03

6 minute read

Hand Gesture Detection - Rev #03

0. Overview

지난 번 Conv3D를 이용한 Post( https://moonlight314.github.io/deep/learning/Hand_Gesture_Detection_Rev_02/ )에서 다룬 방식은 학습이 제대로 이루어 지지 않았습니다.

이번 Post에서는 동영상 각 Frame을 Image처럼 처리하고, Image에 Pre-Trained Model을 이용하여 Feature를 뽑은 다음, 이 Feature들을 RNN에 넣어서 Hand Gesture Detection을 진행해 보도록 하겠습니다.

전체적인 진행 방식은 아래와 같습니다.

Image에서 Feature를 Extract를 할 Pre-Trained Model로는 MobileNetV2를 사용하도록 하겠습니다.

1. Load Module

필요한 Module을 Load합니다.

특별한 Module은 없고, 늘 사용하던 Module들입니다.

import pandas as pd
import cv2 as cv
import numpy as np
from collections import deque

from sklearn.preprocessing import LabelEncoder 
from sklearn.model_selection import train_test_split

import tensorflow as tf
from tensorflow.keras.applications import MobileNetV2
from tensorflow.keras.applications.mobilenet_v2 import preprocess_input
from tensorflow.keras.models import Sequential , Model
from tensorflow.keras.layers import LSTM, Dense, Bidirectional , Dropout , BatchNormalization , GlobalAveragePooling2D

2. Feature Extractor

상수들을 정의합니다.

MAX_FRAME = 128
BATCH_SIZE = 32
EPOCHS = 5
DEST_SIZE = (160,160)

앞서 말했듯이 Feature Extractor로 MobileNetV2를 사용하도록 하겠습니다.

Weight는 Imagenet용으로 초기화하고 Top은 날립니다.

1차원 Shape으로 만들기 위해 AveragePooling을 사용하고, Freeze 시킵니다.

이것으로 Image로 부터 Feature를 뽑아낼 준비는 끝났습니다.

def Feature_Extractor():
    model = Sequential()

    model.add( MobileNetV2(include_top = False, 
                        input_shape=(160, 160, 3),
                        ))   

    model.add(GlobalAveragePooling2D()) 
    model.trainable = False

    return model

3. RNN Model

RNN Model도 앞선 Post에서와 유사하게 만들었습니다.

다만, Input Shape은 Feature Extractor의 Output을 감안하여 변형하였습니다.

def Make_RNN_Model():
    model = Sequential()
    model.add(Bidirectional( LSTM(512,  return_sequences=True , activation='tanh'), input_shape=(MAX_FRAME , 1280)))
    model.add(Bidirectional( LSTM(256,  return_sequences=True , activation='tanh')))
    model.add(Bidirectional( LSTM(64,  return_sequences=False , activation='tanh')))
    model.add(Dense(32, activation='relu'))
    model.add(Dense(4, activation='softmax'))

    return model

4. Generate Train Data

손 동작을 녹화한 Video File로 부터 Feature Extraction하고 모두 동일한 Frame Size로 Length를 맞추는 함수들입니다.

Adjust_Length는 Feature의 Length가 Max Frame과 맞지 않은 경우 앞뒤로 Dummy Data를 붙여서 Length를 모두 동일하도록 해줍니다.

def Adjust_Length( data ):
    
    # Dummy Data로 사용할 Feature
    # 앞쪽은 가장 앞쪽으로 채우고, 뒤쪽은 가장 뒤쪽의 Data로 채운다.
    front = data[0]
    back = data[-1]

    d = np.array(data)
    data = deque(data)

    length = d.shape[0]

    for _ in range(int((MAX_FRAME - length)/2)):
        data.append( back )
        data.appendleft( front )

    if len(data) != MAX_FRAME:
        data.append( back )

    aligned_data = np.array(data)
    
    return aligned_data

generate_train_data 함수는 실제 동영상 File을 읽어서 각 Frame을 Feature Extractor에 넣어 결과를 모은 후, 앞서 설명한 Adjust 과정을 거쳐 RNN Model의 입력에 넣을 수 있도록 Train Data를 생성하는 일을 합니다.

def generate_train_data( file_path , label , feature_extractor):     

    output = []

    for f in file_path:
        
        batch_frames = []

        filename = f.replace("./Train_Data","C:/Users/csyi/Desktop/Hand_Gesture_Detection/Train_Data")
        filename = filename.replace("'","")   

        cap = cv.VideoCapture( filename )

        if cap.isOpened() == False:
            print("Open Error")

        data = deque([])

        while( True ):
            ret, frame = cap.read()

            if ret == False:
                break

            frame = cv.resize( frame, dsize=DEST_SIZE, interpolation=cv.INTER_AREA )
            frame = tf.keras.applications.mobilenet_v2.preprocess_input( frame )
            batch_frames.append( frame )
        
        batch_frames = np.reshape( np.array(batch_frames) , (-1,160,160,3))

        ret = feature_extractor.predict_on_batch( batch_frames )

        data = Adjust_Length( ret )
        output.append(data)
        
    output = np.reshape(np.array(output) , (-1,128,1280))
    label = np.reshape(np.array(label) , (-1,4))

    return output , label

Feature Extractor인 MobileNetV2의 구조입니다.

# Load Feature Extractor
feature_extractor = Feature_Extractor()
feature_extractor.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
mobilenetv2_1.00_160 (Functi (None, 5, 5, 1280)        2257984   
_________________________________________________________________
global_average_pooling2d (Gl (None, 1280)              0         
=================================================================
Total params: 2,257,984
Trainable params: 0
Non-trainable params: 2,257,984
_________________________________________________________________

RNN도 만들어 볼까요?

RNN_Model = Make_RNN_Model()

RNN_Model.compile(  optimizer=tf.keras.optimizers.Adam(1e-3),
                    loss='categorical_crossentropy',
                    metrics=['categorical_accuracy']
)

Meta Data File을 읽어옵니다.

# Load meta data file
meta = pd.read_csv("Meta_Data_220117_Rev_01.csv")

# Data File Path
file_path = meta['file_path'].tolist()

# Label 
labels = meta['action'].tolist()
print(len(file_path) , len(labels))

389 389

Label도 준비합니다.

le = LabelEncoder()
le_action = le.fit(labels)
le_action = le.transform(labels)
print(le.classes_)

y = tf.keras.utils.to_categorical(le_action, num_classes=4)
print(y)

['1_Finger_Click' '2_Fingers_Left' '2_Fingers_Right' 'Shake_Hand']
[[1. 0. 0. 0.]
 [1. 0. 0. 0.]
 [1. 0. 0. 0.]
 ...
 [0. 0. 0. 1.]
 [0. 0. 0. 1.]
 [0. 0. 0. 1.]]

# Train / Test 나누기
X_train, X_test, y_train, y_test = train_test_split(file_path, y, test_size=0.25 , stratify = y)

tmp_val_data = []
tmp_val_target = []

대망의 Training Stage입니다.

Feature Extractor에서 나온 값을 RNN의 Input으로 넣는 과정으로 진행됩니다.

for epoch in range(EPOCHS):

    print("### Epoch : {0} ###\n\n".format(epoch))

    for idx in range( 0 , len(X_train) , BATCH_SIZE):

        tmp_train_data = []
        tmp_target = []

        batch_file_list = []
        batch_target = []

        for batch in range(BATCH_SIZE):                

            if idx+batch >= len(X_train):
                break

            batch_file_list.append( X_train[idx+batch] )
            batch_target.append( y_train[idx+batch] )

        train_data , target = generate_train_data( batch_file_list, batch_target , feature_extractor)

        RNN_Model.fit(  x = train_data,
                        y = target,            
                        verbose=1)

    # Eval.
    print("### Evaluation... Epoch : {0} ###".format(epoch))

    if len(tmp_val_data) == 0:
        tmp_val_data , tmp_val_target = generate_train_data( X_test , y_test , feature_extractor)

    ret = RNN_Model.evaluate(tmp_val_data , tmp_val_target )

### Epoch : 0 ###


1/1 [==============================] - 6s 6s/step - loss: 1.3582 - categorical_accuracy: 0.2188
1/1 [==============================] - 0s 246ms/step - loss: 1.6103 - categorical_accuracy: 0.2812
1/1 [==============================] - 0s 248ms/step - loss: 1.9793 - categorical_accuracy: 0.0938
1/1 [==============================] - 0s 246ms/step - loss: 1.2979 - categorical_accuracy: 0.4062
1/1 [==============================] - 0s 244ms/step - loss: 1.4777 - categorical_accuracy: 0.1875
1/1 [==============================] - 0s 247ms/step - loss: 1.3085 - categorical_accuracy: 0.2812
1/1 [==============================] - 0s 247ms/step - loss: 1.3324 - categorical_accuracy: 0.2188
1/1 [==============================] - 0s 247ms/step - loss: 1.2749 - categorical_accuracy: 0.4062
1/1 [==============================] - 0s 244ms/step - loss: 1.1534 - categorical_accuracy: 0.4375
1/1 [==============================] - 4s 4s/step - loss: 1.1634 - categorical_accuracy: 0.0000e+00
### Evaluation... Epoch : 0 ###
4/4 [==============================] - 2s 89ms/step - loss: 1.2826 - categorical_accuracy: 0.4592
### Epoch : 1 ###


1/1 [==============================] - 0s 246ms/step - loss: 1.2443 - categorical_accuracy: 0.4375
1/1 [==============================] - 0s 245ms/step - loss: 1.1586 - categorical_accuracy: 0.4688
1/1 [==============================] - 0s 243ms/step - loss: 1.0803 - categorical_accuracy: 0.6562
1/1 [==============================] - 0s 246ms/step - loss: 0.9259 - categorical_accuracy: 0.7812
1/1 [==============================] - 0s 243ms/step - loss: 0.9260 - categorical_accuracy: 0.5938
1/1 [==============================] - 0s 240ms/step - loss: 0.7924 - categorical_accuracy: 0.7500
1/1 [==============================] - 0s 244ms/step - loss: 0.6781 - categorical_accuracy: 0.8125
1/1 [==============================] - 0s 247ms/step - loss: 0.5861 - categorical_accuracy: 0.8438
1/1 [==============================] - 0s 238ms/step - loss: 0.5703 - categorical_accuracy: 0.8438
1/1 [==============================] - 0s 106ms/step - loss: 0.5877 - categorical_accuracy: 1.0000
### Evaluation... Epoch : 1 ###
4/4 [==============================] - 0s 88ms/step - loss: 0.8222 - categorical_accuracy: 0.6020
### Epoch : 2 ###


1/1 [==============================] - 0s 240ms/step - loss: 0.7703 - categorical_accuracy: 0.5938
1/1 [==============================] - 0s 239ms/step - loss: 0.6549 - categorical_accuracy: 0.8438
1/1 [==============================] - 0s 245ms/step - loss: 0.6331 - categorical_accuracy: 0.8438
1/1 [==============================] - 0s 242ms/step - loss: 0.5142 - categorical_accuracy: 0.8125
1/1 [==============================] - 0s 228ms/step - loss: 0.4535 - categorical_accuracy: 0.7812
1/1 [==============================] - 0s 251ms/step - loss: 0.3349 - categorical_accuracy: 0.8438
1/1 [==============================] - 0s 244ms/step - loss: 0.2939 - categorical_accuracy: 1.0000
1/1 [==============================] - 0s 243ms/step - loss: 0.3022 - categorical_accuracy: 1.0000
1/1 [==============================] - 0s 242ms/step - loss: 0.2264 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 107ms/step - loss: 0.1069 - categorical_accuracy: 1.0000
### Evaluation... Epoch : 2 ###
4/4 [==============================] - 0s 87ms/step - loss: 0.2742 - categorical_accuracy: 0.9184
### Epoch : 3 ###


1/1 [==============================] - 0s 249ms/step - loss: 0.2286 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 248ms/step - loss: 0.2914 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 241ms/step - loss: 0.2398 - categorical_accuracy: 1.0000
1/1 [==============================] - 0s 247ms/step - loss: 0.1969 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 242ms/step - loss: 0.1515 - categorical_accuracy: 1.0000
1/1 [==============================] - 0s 245ms/step - loss: 0.2328 - categorical_accuracy: 0.8438
1/1 [==============================] - 0s 248ms/step - loss: 0.1201 - categorical_accuracy: 1.0000
1/1 [==============================] - 0s 247ms/step - loss: 0.2063 - categorical_accuracy: 0.9375
1/1 [==============================] - 0s 241ms/step - loss: 0.0878 - categorical_accuracy: 1.0000
1/1 [==============================] - 0s 105ms/step - loss: 0.0248 - categorical_accuracy: 1.0000
### Evaluation... Epoch : 3 ###
4/4 [==============================] - 0s 87ms/step - loss: 0.1733 - categorical_accuracy: 0.9490
### Epoch : 4 ###


1/1 [==============================] - 0s 241ms/step - loss: 0.1172 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 243ms/step - loss: 0.1761 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 250ms/step - loss: 0.2952 - categorical_accuracy: 0.8750
1/1 [==============================] - 0s 244ms/step - loss: 0.1936 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 243ms/step - loss: 0.3419 - categorical_accuracy: 0.9062
1/1 [==============================] - 0s 244ms/step - loss: 0.1645 - categorical_accuracy: 0.9375
1/1 [==============================] - 0s 239ms/step - loss: 0.3557 - categorical_accuracy: 0.8438
1/1 [==============================] - 0s 253ms/step - loss: 0.1632 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 243ms/step - loss: 0.1602 - categorical_accuracy: 0.9688
1/1 [==============================] - 0s 107ms/step - loss: 0.0483 - categorical_accuracy: 1.0000
### Evaluation... Epoch : 4 ###
4/4 [==============================] - 0s 86ms/step - loss: 0.4784 - categorical_accuracy: 0.8469

다행히도 Training이 되네요. Train / Validation Data 모두에서 Loss / Acc.가 좋아지는 모습을 볼 수 있습니다.

Share on

Twitter Facebook LinkedIn

MoonLight

Hand Gesture Detection - Rev #03

Hand Gesture Detection - Rev #03

0. Overview

1. Load Module

2. Feature Extractor

3. RNN Model

4. Generate Train Data

Share on

You May Also Enjoy

대규모 머신러닝 시스템 디자인 패턴(Distributed Machine Learning Patterns)

소프트웨어 엔지니어 가이드북(The Software Engineer’s Guidebook)

이펙티브 러스트(Effective Rust)

세상에서 제일 쉬운 챗GPT 프롬프트 엔지니어링