前言
手语视频流的识别有两种大的分类，一种是直接将视频输入进网络，一种是识别了关键点之后再进入网络。所以这篇文章我就要来讲讲如何用mediapipe对手语视频进行关键点标注。
代码
需要直接使用代码的，我就放这里了。环境自己配置一下吧，不太记得了。
python
import os
import cv2
import numpy as np
import mediapipe as mp
from concurrent.futures import ThreadPoolExecutor
# 关键点过滤设置
filtered_hand = list(range(21))
filtered_pose = [11, 12, 13, 14, 15, 16] # 只保留躯干和手臂关键点
HAND_NUM = len(filtered_hand)
POSE_NUM = len(filtered_pose)
# 初始化MediaPipe模型（增加检测参数）
mp_hands = mp.solutions.hands
mp_pose = mp.solutions.pose
hands = mp_hands.Hands(
static_image_mode=False,
max_num_hands=2,
min_detection_confidence=0.1,#太高的话，没识别到就不识别，比较低能识别的比较全（没有干扰的情况下低比较好）
min_tracking_confidence=0.1#太高，没追踪到也会放弃，比较低的连续性会比较好
)
pose = mp_pose.Pose(
static_image_mode=False,
model_complexity=1,
min_detection_confidence=0.7,
min_tracking_confidence=0.5
)
def get_frame_landmarks(frame):
"""获取单帧关键点（修复线程安全问题）"""
all_landmarks = np.full((HAND_NUM * 2 + POSE_NUM, 3), np.nan) # 初始化为NaN
# 改为顺序执行确保数据可靠性
# 手部关键点
results_hands = hands.process(frame)
if results_hands.multi_hand_landmarks:
for i, hand_landmarks in enumerate(results_hands.multi_hand_landmarks[:2]): # 最多两只手
hand_type = results_hands.multi_handedness[i].classification[0].index
points = np.array([(lm.x, lm.y, lm.z) for lm in hand_landmarks.landmark])
if hand_type == 0: # 右手
all_landmarks[:HAND_NUM] = points
else: # 左手
all_landmarks[HAND_NUM:HAND_NUM * 2] = points
# 身体关键点
results_pose = pose.process(frame)
if results_pose.pose_landmarks:
pose_points = np.array([(lm.x, lm.y, lm.z) for lm in results_pose.pose_landmarks.landmark])
all_landmarks[HAND_NUM * 2:HAND_NUM * 2 + POSE_NUM] = pose_points[filtered_pose]
return all_landmarks
def get_video_landmarks(video_path, start_frame=1, end_frame=-1):
"""获取视频关键点（添加调试信息）"""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
print(f"无法打开视频文件: {video_path}")
return np.empty((0, HAND_NUM * 2 + POSE_NUM, 3))
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
if end_frame < 0 or end_frame > total_frames:
end_frame = total_frames
valid_frames = []
frame_index = 0
while cap.isOpened():
ret, frame = cap.read()
if not ret or frame_index > end_frame:
break
if frame_index >= start_frame:
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
landmarks = get_frame_landmarks(frame_rgb)
# 检查是否检测到有效关键点
if not np.all(np.isnan(landmarks)):
valid_frames.append(landmarks)
else:
print(f"第 {frame_index} 帧未检测到关键点")
frame_index += 1
cap.release()
if not valid_frames:
print("警告：未检测到任何关键点")
return np.empty((0, HAND_NUM * 2 + POSE_NUM, 3))
return np.stack(valid_frames)
def draw_landmarks(video_path, output_path, landmarks):
"""绘制关键点到视频"""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
print(f"无法打开视频文件: {video_path}")
return
fps = int(cap.get(cv2.CAP_PROP_FPS))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
landmark_index = 0
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
if landmark_index < len(landmarks):
# 绘制关键点
for i, (x, y, _) in enumerate(landmarks[landmark_index]):
if not np.isnan(x) and not np.isnan(y):
px, py = int(x * width), int(y * height)
# 右手绿色，左手红色，身体蓝色
color = (0, 255, 0) if i < HAND_NUM else \
(0, 0, 255) if i < HAND_NUM * 2 else \
(255, 0, 0)
cv2.circle(frame, (px, py), 4, color, -1)
landmark_index += 1
out.write(frame)
cap.release()
out.release()
# 处理所有视频
video_root = "./doc/补充版/正式数据集/"
output_root = "./doc/save/"
if not os.path.exists(output_root):
os.makedirs(output_root)
for video_name in os.listdir(video_root):
if not video_name.endswith(('.mp4', '.avi', '.mov')):
continue
video_path = os.path.join(video_root, video_name)
print(f"\n处理视频: {video_name}")
# 获取关键点
landmarks = get_video_landmarks(video_path)
print(f"获取到 {len(landmarks)} 帧关键点")
# 保存npy文件
base_name = os.path.splitext(video_name)[0]
np.save(os.path.join(output_root,"npy", f"{base_name}.npy"), landmarks)
# 生成带关键点的视频
output_video = os.path.join(output_root, "MP4",f"{base_name}_landmarks.mp4")
draw_landmarks(video_path, output_video, landmarks)
print("全部处理完成！")使用比较简单，修改video_root为视频目录路径，output_root为结果输出目录路径就可以正常使用了！
前置处理
python
# 关键点过滤设置
filtered_hand = list(range(21))
filtered_pose = [11, 12, 13, 14, 15, 16] # 只保留躯干和手臂关键点
HAND_NUM = len(filtered_hand)
POSE_NUM = len(filtered_pose)
)这里需要选取你需要的关键点，手部正常来说每个手21个，姿态和脸部的关键点也可以自己选择保留什么，网上可以查到每个点对应数字。
python
# 初始化MediaPipe模型（增加检测参数）
mp_hands = mp.solutions.hands
mp_pose = mp.solutions.pose
hands = mp_hands.Hands(
static_image_mode=False,
max_num_hands=2,
min_detection_confidence=0.1,#太高的话，没识别到就不识别，比较低能识别的比较全（没有干扰的情况下低比较好）
min_tracking_confidence=0.1#太高，没追踪到也会放弃，比较低的连续性会比较好
)
pose = mp_pose.Pose(
static_image_mode=False,
model_complexity=1,
min_detection_confidence=0.7,
min_tracking_confidence=0.5参数调整，对于手部和姿态都可以进行单独的参数调整，static_image_mode是是否是图片，False代表不是，我这里是视频，如果是视频的话，后面就还有一个min_tracking_confidence追踪阈值，而图片不具有时间连续性，所以用不到这个参数。max_num_hands是最大会识别到有几个手，后面两个参数我也写了怎么调。姿态参数基本同理，有一些区别可以自己查一下。
函数讲解
python
def get_frame_landmarks(frame):
"""获取单帧关键点（修复线程安全问题）"""
all_landmarks = np.full((HAND_NUM * 2 + POSE_NUM, 3), np.nan) # 初始化为NaN
# 改为顺序执行确保数据可靠性
# 手部关键点
results_hands = hands.process(frame)
if results_hands.multi_hand_landmarks:
for i, hand_landmarks in enumerate(results_hands.multi_hand_landmarks[:2]): # 最多两只手
hand_type = results_hands.multi_handedness[i].classification[0].index
points = np.array([(lm.x, lm.y, lm.z) for lm in hand_landmarks.landmark])
if hand_type == 0: # 右手
all_landmarks[:HAND_NUM] = points
else: # 左手
all_landmarks[HAND_NUM:HAND_NUM * 2] = points
# 身体关键点
results_pose = pose.process(frame)
if results_pose.pose_landmarks:
pose_points = np.array([(lm.x, lm.y, lm.z) for lm in results_pose.pose_landmarks.landmark])
all_landmarks[HAND_NUM * 2:HAND_NUM * 2 + POSE_NUM] = pose_points[filtered_pose]
return all_landmarks对于单帧进行处理，先对所有关键点留np的位置，全部填充NaN,再分别对手部关键点和肢体关键点进行识别，将识别的点填入原先的数组里面，得到最后要返回的关键点数组。
python
def get_video_landmarks(video_path, start_frame=1, end_frame=-1):
"""获取视频关键点（添加调试信息）"""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
print(f"无法打开视频文件: {video_path}")
return np.empty((0, HAND_NUM * 2 + POSE_NUM, 3))
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
if end_frame < 0 or end_frame > total_frames:
end_frame = total_frames
valid_frames = []
frame_index = 0
while cap.isOpened():
ret, frame = cap.read()
if not ret or frame_index > end_frame:
break
if frame_index >= start_frame:
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
landmarks = get_frame_landmarks(frame_rgb)
# 检查是否检测到有效关键点
if not np.all(np.isnan(landmarks)):
valid_frames.append(landmarks)
else:
print(f"第 {frame_index} 帧未检测到关键点")
frame_index += 1
cap.release()
if not valid_frames:
print("警告：未检测到任何关键点")
return np.empty((0, HAND_NUM * 2 + POSE_NUM, 3))
return np.stack(valid_frames)处理视频帧的关键点识别，读取视频的每一帧，分别做通道BGR转RGB和调用单帧处理函数对其进行处理，将每一帧的结果堆叠起来返回。
python
def draw_landmarks(video_path, output_path, landmarks):
"""绘制关键点到视频"""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
print(f"无法打开视频文件: {video_path}")
return
fps = int(cap.get(cv2.CAP_PROP_FPS))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
landmark_index = 0
while cap.isOpened():
ret, frame = cap.read()
if not ret:
break
if landmark_index < len(landmarks):
# 绘制关键点
for i, (x, y, _) in enumerate(landmarks[landmark_index]):
if not np.isnan(x) and not np.isnan(y):
px, py = int(x * width), int(y * height)
# 右手绿色，左手红色，身体蓝色
color = (0, 255, 0) if i < HAND_NUM else \
(0, 0, 255) if i < HAND_NUM * 2 else \
(255, 0, 0)
cv2.circle(frame, (px, py), 4, color, -1)
landmark_index += 1
out.write(frame)
cap.release()
out.release()绘制结果关键点函数，将视频路径和输出路径以及识别的关键点数组传入，读取视频，对每一帧的图片每一个关键点进行绘制，画圈圈，然后将帧写入保存。
总结
整个路线还是比较清晰的，由于我使用的数据视频背景比较简单，不太会出现误识别，所以我的参数调的很低，但是不知道为什么还是会出现掉帧的情况，需要后续研究一下。