首先我们将采集的视频文件解码抽取关键帧，每隔4帧保存一次，然后对图像进行中心裁剪和预处理，代码如下：

def load_video(file_name):
    cap = cv2.VideoCapture(file_name) 
    # 每隔多少帧抽取一次
    frame_interval = 4
    frames = []
    count = 0
    while True:
        ret, frame = cap.read()
        if not ret:
            break
        
        # 每隔frame_interval帧保存一次
        if count % frame_interval == 0:
            # 中心裁剪    
            frame = crop_center_square(frame)
            # 缩放
            frame = cv2.resize(frame, (IMG_SIZE, IMG_SIZE))
            # BGR -> RGB  [0,1,2] -> [2,1,0]
            frame = frame[:, :, [2, 1, 0]]
            frames.append(frame)
        count += 1
        
    return np.array(frames)

然后我们创建图像特征提取器，使用预训练模型InceptionResNetV2提取图像特征，代码如下：

def get_feature_extractor():
    feature_extractor = keras.applications.inception_resnet_v2.InceptionResNetV2(
        weights = 'imagenet',
        include_top = False,
        pooling = 'avg',
        input_shape = (IMG_SIZE, IMG_SIZE, 3)
    )
    
    preprocess_input = keras.applications.inception_resnet_v2.preprocess_input
    
    inputs = keras.Input((IMG_SIZE, IMG_SIZE, 3))
    preprocessed = preprocess_input(inputs)
    outputs = feature_extractor(preprocessed)
    
    model = keras.Model(inputs, outputs, name = 'feature_extractor')
    
    return model

接着提取视频特征向量，如果视频不足40帧就创建全0数组进行补白：

def load_data(videos, labels):
    
    video_features = []

    for video in tqdm(videos):
        frames = load_video(video)
        counts = len(frames)
        # 如果帧数小于MAX_SEQUENCE_LENGTH
        if counts < MAX_SEQUENCE_LENGTH:
            # 补白
            diff = MAX_SEQUENCE_LENGTH - counts
            # 创建全0的numpy数组
            padding = np.zeros((diff, IMG_SIZE, IMG_SIZE, 3))
            # 数组拼接
            frames = np.concatenate((frames, padding))
        # 获取前MAX_SEQUENCE_LENGTH帧画面
        frames = frames[:MAX_SEQUENCE_LENGTH, :]
        # 批量提取特征
        video_feature = feature_extractor.predict(frames)
        video_features.append(video_feature)
        
    return np.array(video_features), np.array(labels)

最后创建VIT Model，代码如下：

# 位置编码
class PositionalEmbedding(layers.Layer):
    def __init__(self, seq_length, output_dim):
        super().__init__()
        # 构造从0~MAX_SEQUENCE_LENGTH的列表
        self.positions = tf.range(0, limit=MAX_SEQUENCE_LENGTH)
        self.positional_embedding = layers.Embedding(input_dim=seq_length, output_dim=output_dim)
    
    def call(self,x):
        # 位置编码
        positions_embedding = self.positional_embedding(self.positions)
        # 输入相加
        return x + positions_embedding

# 编码器
class TransformerEncoder(layers.Layer):
    
    def __init__(self, num_heads, embed_dim):
        super().__init__()
        self.p_embedding = PositionalEmbedding(MAX_SEQUENCE_LENGTH, NUM_FEATURES)
        self.attention = layers.MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim, dropout=0.1)
        self.layernorm = layers.LayerNormalization()
    
    def call(self,x):
        # positional embedding
        positional_embedding = self.p_embedding(x)
        # self attention
        attention_out = self.attention(
            query = positional_embedding,
            value = positional_embedding,
            key = positional_embedding,
            attention_mask = None
        )
        # layer norm with residual connection        
        output = self.layernorm(positional_embedding + attention_out)
        return output

def video_cls_model(class_vocab):
    # 类别数量
    classes_num = len(class_vocab)
    # 定义模型
    model = keras.Sequential([
        layers.InputLayer(input_shape=(MAX_SEQUENCE_LENGTH, NUM_FEATURES)),
        TransformerEncoder(2, NUM_FEATURES),
        layers.GlobalMaxPooling1D(),
        layers.Dropout(0.1),
        layers.Dense(classes_num, activation="softmax")
    ])
    # 编译模型
    model.compile(optimizer = keras.optimizers.Adam(1e-5), 
                  loss = keras.losses.SparseCategoricalCrossentropy(from_logits=False),
                  metrics = ['accuracy']
    )
    return model

模型训练

完整体验可以点击Run in ModelArts一键运行我发布的Notebook：

最终模型在整个数据集上的准确率达到87%，即在小数据集上训练取得了较为不错的结果。

视频推理

首先加载VIT Model，获取视频类别索引标签：

import random
# 加载模型
model = tf.keras.models.load_model('saved_model')
# 类别标签
label_to_name = {0:'无效手势', 1:'上滑', 2:'下滑', 3:'左滑', 4:'右滑', 5:'打开', 6:'关闭', 7:'放大', 8:'缩小'}

然后使用图像特征提取器InceptionResNetV2提取视频特征：

# 获取视频特征
def getVideoFeat(frames):
    
    frames_count = len(frames)
    
    # 如果帧数小于MAX_SEQUENCE_LENGTH
    if frames_count < MAX_SEQUENCE_LENGTH:
        # 补白
        diff = MAX_SEQUENCE_LENGTH - frames_count
        # 创建全0的numpy数组
        padding = np.zeros((diff, IMG_SIZE, IMG_SIZE, 3))
        # 数组拼接
        frames = np.concatenate((frames, padding))

    # 取前MAX_SEQ_LENGTH帧
    frames = frames[:MAX_SEQUENCE_LENGTH,:]
    # 计算视频特征 N, 1536
    video_feat = feature_extractor.predict(frames)

    return video_feat

最后将视频序列的特征向量输入Transformer Encoder进行预测：

# 视频预测
def testVideo():
    test_file = random.sample(videos, 1)[0]
    label = test_file.split('_')[-2]

    print('文件名:{}'.format(test_file) )
    print('真实类别:{}'.format(label_to_name.get(int(label))) )

    # 读取视频每一帧
    frames = load_video(test_file)
    # 挑选前帧MAX_SEQUENCE_LENGTH显示
    frames = frames[:MAX_SEQUENCE_LENGTH].astype(np.uint8)
    # 保存为GIF
    imageio.mimsave('animation.gif', frames, duration=10)
    # 获取特征
    feat = getVideoFeat(frames)
    # 模型推理
    prob = model.predict(tf.expand_dims(feat, axis=0))[0]
    
    print('预测类别：')
    for i in np.argsort(prob)[::-1][:5]:
        print('{}: {}%'.format(label_to_name[i], round(prob[i]*100, 2)))
    
    return display(Image(open('animation.gif', 'rb').read()))

模型预测结果：

文件名:hand_gesture/woman_014_0_7.mp4
真实类别:无效手势
预测类别：
无效手势: 99.82%
下滑: 0.12%
关闭: 0.04%
左滑: 0.01%
打开: 0.01%

点击关注，第一时间了解华为云新鲜技术~