阅读量:2
优化人物淡入淡出时，人脸判断依然有效的情况：
引入sample_rate参数来控制帧率采样，减少不必要的处理。 使用min_confidence参数来过滤低置信度的人脸检测结果。 使用滑动窗口计算置信度的移动平均，以检测淡出效果。fade_threshold参数用于决定何时认为人脸真正淡出。 我们还增加了timeout参数，用于在没有人脸被检测到一定时间后停止搜索。 注意，face_recognition库的face_encodings方法返回的编码并不直接对应置信度，但我们可以利用编码的范数作为置信度的代理。这个范数通常在人脸存在时会相对较小，不存在时较大，因此可以作为置信度的反向指标。 这个版本的代码应该能够更准确地处理淡入淡出效果，并在保持较高检测灵敏度的同时，加快处理速度。不过，你可能需要根据实际情况调整sample_rate、min_confidence、fade_threshold和timeout等参数。 完整代码：
# python data_utils/pre_video/cut_crop_fps.py import cv2 import math import numpy as np import face_recognition from moviepy.editor import VideoFileClip, concatenate_videoclips from tqdm import tqdm def find_host_face_location(video_path): """ 在视频的前几秒内检测并返回主持人面部的大致位置 """ cap = cv2.VideoCapture(video_path) found_face = False host_face_location = None while cap.isOpened(): ret, frame = cap.read() if not ret: break # 缩小帧尺寸以加快处理速度 small_frame = cv2.resize(frame, (0, 0), fx=0.25, fy=0.25) rgb_small_frame = small_frame[:, :, ::-1] # 检测人脸 face_locations = face_recognition.face_locations(rgb_small_frame) if face_locations: # 取第一张脸的位置，假设主持人位于视频画面的中心位置附近 host_face_location = face_locations[0] # 将位置放大回原始大小 host_face_location = (host_face_location[0]*4, host_face_location[1]*4, host_face_location[2]*4, host_face_location[3]*4) found_face = True break cap.release() return host_face_location if found_face else None def calculate_cropping_box(face_location, frame_shape): """ 根据主持人面部位置计算裁剪框 """ top, right, bottom, left = face_location center_x, center_y = (left + right) // 2, (top + bottom) // 2 half_width, half_height = 256, 256 left_cropped = max(center_x - half_width, 0) top_cropped = max(center_y - half_height, 0) right_cropped = min(center_x + half_width, frame_shape[1]) bottom_cropped = min(center_y + half_height, frame_shape[0]) return (top_cropped, right_cropped, bottom_cropped, left_cropped) def find_first_last_face(video_path, sample_rate=3, min_confidence=0.5, fade_threshold=0.5, timeout=2.0): """ 找到视频中第一次和最后一次出现人脸的时间戳，优化处理速度和淡出效果 """ cap = cv2.VideoCapture(video_path) fps = cap.get(cv2.CAP_PROP_FPS) first_face_time = None last_face_time = 0 last_face_detected_time = 0 timeout_counter = 0.0 face_confidences = [] timestamps = [] while cap.isOpened(): ret, frame = cap.read() if not ret: break # 按照sample_rate进行帧率采样 if len(timestamps) > 0 and (len(timestamps) % sample_rate) != 0: continue timestamp = cap.get(cv2.CAP_PROP_POS_MSEC) / 1000 # Convert to seconds # 缩小帧尺寸以加快处理速度 small_frame = cv2.resize(frame, (0, 0), fx=0.25, fy=0.25) rgb_small_frame = small_frame[:, :, ::-1] # 检测人脸及其置信度 face_locations = face_recognition.face_locations(rgb_small_frame, model='hog') # 可以选择 'cnn' 或 'hog' face_encodings = face_recognition.face_encodings(rgb_small_frame, face_locations) face_confidence_scores = [np.linalg.norm(face_encoding) for face_encoding in face_encodings] if face_locations: if not first_face_time: first_face_time = timestamp last_face_detected_time = timestamp timeout_counter = 0.0 # 计算平均置信度 avg_confidence = sum(face_confidence_scores) / len(face_confidence_scores) face_confidences.append(avg_confidence) timestamps.append(timestamp) # 检查是否低于置信度阈值 if avg_confidence < min_confidence: face_confidences[-1] = 0 else: timeout_counter += 1/fps face_confidences.append(0) timestamps.append(timestamp) if timeout_counter > timeout: last_face_time = last_face_detected_time break cap.release() # 后处理逻辑：滑动窗口检测强度 window_size = int(timeout * fps / sample_rate) if window_size > 1: moving_averages = np.convolve(face_confidences, np.ones(window_size)/window_size, mode='valid') moving_timestamps = timestamps[window_size-1:] for i, avg in enumerate(moving_averages): if avg < fade_threshold and last_face_time == 0: last_face_time = moving_timestamps[i] break # 如果整个视频都没有检测到人脸，设置last_face_time为None if last_face_time == 0: last_face_time = None return first_face_time, last_face_time def process_video(input_path, output_path): """ 处理视频，裁剪并调整帧率 """ # 检测主持人面部位置 host_face_location = find_host_face_location(input_path) if host_face_location is None: print(f"No face detected in video {input_path}") return # 读取视频，获取视频的宽度和高度 clip = VideoFileClip(input_path) frame_shape = clip.size[::-1] # 电影剪辑的尺寸是(width, height)，我们需要(height, width) # 计算裁剪框 cropping_box = calculate_cropping_box(host_face_location, frame_shape) # 找到第一次和最后一次出现人脸的时间 first_face_time, last_face_time = find_first_last_face(input_path) print(f"First face time: {first_face_time}, Last face time: {last_face_time}") # 裁剪视频以保留第一次和最后一次出现人脸的部分 start_trim = math.ceil(first_face_time) # 向上取整 end_trim = math.floor(last_face_time) # 向下取整 print(f"Start trim: {start_trim}, End trim: {end_trim}") trimmed_clip = clip.subclip(start_trim, end_trim) # 裁剪视频 cropped_clip = trimmed_clip.crop(x1=cropping_box[3], y1=cropping_box[0], x2=cropping_box[1], y2=cropping_box[2]) cropped_clip = cropped_clip.resize((512, 512)) # 调整帧率 cropped_clip = cropped_clip.set_fps(25) # 保存最终视频 cropped_clip.write_videofile(output_path, codec='libx264', audio_codec='aac') # 清理资源 cropped_clip.close() if __name__ == "__main__": for i in tqdm(range(1, 76), desc="Processing videos"): print("处理第", i, "个视频") input_path = f"data/dataset/{i}/{i}.mp4" output_path = f"data/dataset/{i}/{i}_fcc.mp4" process_video(input_path, output_path) 