一、多线程优化
1、多线程并行图像拼接，视频数据流检测并将每节车厢的帧数保存，然后直接使用线程调用拼接函数，不影响主程序视频流的检测。
2、基于get_img_asyncio.py程序中修改run.py
python
def run(self, video_num=0):
# 基本数据初始化
run_state = False
img_list = []
cg_list = [[], []]
cut = []
img_num_list = []
begin_num = 0
total_res = []
det_res_list = []
# 线程池相关变量
stich_futures = [] # 存储所有拼接任务的Future对象
stich_tasks = [] # 存储任务元数据
try:
cap = cv2.VideoCapture(self.source, cv2.CAP_FFMPEG)
if not cap.isOpened():
logging.error(f"{self.task_id} 无法打开视频源: {self.source}")
self.running = False
return total_res
cap_info = self.get_cap_info(cap)
self.total_stich_side.basic_shape = cap_info['shape']
self.total_stich_side.cut_shape = [self.center_box[3], self.center_box[2]-self.center_box[0]]
self.total_stich_side.x_area = [self.center_box[0], self.center_box[2]]
self.total_stich_side.basic_num = self.train_num_base
self.total_stich_side.type = self.type
self.total_stich_up.basic_shape = cap_info['shape']
self.total_stich_up.cut_shape = [self.center_box[3], self.center_box[2]-self.center_box[0]]
self.total_stich_up.x_area = [self.center_box[0], self.center_box[2]]
self.total_stich_up.basic_num = self.train_num_base
self.total_stich_up.type = self.type
self.image_shape = cap_info['shape']
pre_dedtect = {'stich': False, 'hitnum': 0}
end_dedtect = 0
logging.info(f"{self.task_id} 成功打开视频源: {self.source}")
stop = 0
start_det = 0
# 创建线程池（根据需求调整max_workers）
thread_pool = ThreadPoolExecutor(max_workers=4, thread_name_prefix=f"Sticher-{self.task_id}")
for total_num in range(cap_info['frames']):
# 检查已完成的拼接任务
completed_futures = []
for future in stich_futures:
if future.done():
try:
res = future.result()
if res is not None:
total_res.append(res)
except Exception as e:
logging.error(f"拼接任务出错: {str(e)}")
traceback.print_exc()
completed_futures.append(future)
# 移除已完成的任务
for future in completed_futures:
stich_futures.remove(future)
ret = cap.grab()
if not ret:
break
if total_num % 100 == 0:
logging.info(f'{self.task_id} total_num:{total_num} ')
ret, frame = cap.retrieve()
if pre_dedtect['hitnum'] == 0 and total_num % 5 != 0:
continue
det_res = self.det.infer(frame)
# 第一节车厢拼接，获取开头相应位置以及初始化必要信息
if not run_state:
if (video_num == 0 and total_num < 100) or len(det_res[1]) == 0:
continue
else:
if not pre_dedtect['stich']:
if self.flitter_railways(det_res):
pre_dedtect['hitnum'] += 1
if pre_dedtect['hitnum'] > 5:
pre_dedtect['stich'] = True
else:
if pre_dedtect['hitnum'] > 0:
pre_dedtect['hitnum'] -= 1
else:
standart_H, intersection = self.calculate_union_and_check_intersection(
det_res[0], det_res[1], self.center_box, run_state)
if not intersection:
continue
begin_num = total_num
run_state = True
else:
standart_H, intersection = self.calculate_union_and_check_intersection(
det_res[0], det_res[1], self.center_box, run_state)
if not intersection:
if len(cg_list[0]) == 0:
stop += 1
if stop >= 15:
break
continue
new_list = [int(abs(x - (self.center_box[2] - self.center_box[0]) / 2)) for x in cg_list[0][0]]
if len(new_list) == 0:
if end_dedtect < 10:
end_dedtect += 1
continue
else:
break
id = new_list.index(min(new_list))
end = img_num_list.index(cg_list[1][id])
cut.append([end, cg_list[0][id]])
if self.total_stich is None:
self.total_stich = self.total_stich_side
# 创建独立的任务数据副本
task_imgs = img_list[:end + 1].copy()
task_dets = det_res_list[:end + 1].copy()
task_cuts = cut.copy()
# 创建新的拼接器实例避免状态冲突
if self.type == 'up':
task_sticher = self.total_stich_up
else:
task_sticher = self.total_stich_side
# 创建拼接器副本（浅拷贝）
task_sticher = copy.copy(task_sticher)
# 设置任务专属参数
init_params = {
'basic_shape': task_sticher.basic_shape,
'cut_shape': task_sticher.cut_shape,
'x_area': task_sticher.x_area,
'save_path': task_sticher.save_path,
'train_num': task_sticher.train_num_base + len(stich_tasks),
'type': task_sticher.type
}
# 延迟初始化任务专属状态
if hasattr(task_sticher, 'initialize'):
task_sticher.initialize(init_params)
# 提交任务到线程池
future = thread_pool.submit(
stich_worker_thread,
task_imgs,
task_dets,
task_cuts,
task_sticher
)
stich_futures.append(future)
stich_tasks.append({
'end_index': end,
'train_num': self.train_num_base + len(stich_tasks)
})
# 更新数据结构
img_list = img_list[end:]
det_res_list = det_res_list[end:]
img_num_list = img_num_list[end:]
cut = [[0, cut[-1][1]]]
cg_list = [[], []]
self.begin_cut = True
else:
det_box, cg = self.calculate_intersection_of_union_with_input(det_res, self.center_box)
img_list.append(frame)
img_num_list.append(total_num)
det_res_list.append(det_res)
if cg[0] > 0:
if len(cut) > 0:
if total_num - img_num_list[0] < 30:
continue
cg_list[0].append(cg)
cg_list[1].append(total_num)
else:
if len(cg_list[0]) != 0:
new_list = [int(abs(x[0] - (self.center_box[2] - self.center_box[0]) / 2)) for x in cg_list[0]]
id = new_list.index(min(new_list))
if (total_num - begin_num) - len(cg_list[0]) < 50 and not self.begin_cut:
logging.info(f"{self.task_id}开始裁切起始位置")
begin = img_num_list.index(cg_list[1][id])
img_list = img_list[begin:]
det_res_list = det_res_list[begin:]
img_num_list = img_num_list[begin:]
cut.append([begin, cg_list[0][id]])
cg_list = [[], []]
self.begin_cut = True
else:
end = img_num_list.index(cg_list[1][id])
cut.append([end, cg_list[0][id]])
if self.total_stich is None:
self.total_stich = self.total_stich_side
# 创建独立的任务数据副本
task_imgs = img_list[:end + 1].copy()
task_dets = det_res_list[:end + 1].copy()
task_cuts = cut.copy()
# 创建新的拼接器实例
if self.type == 'up':
task_sticher = self.total_stich_up
else:
task_sticher = self.total_stich_side
task_sticher = copy.copy(task_sticher)
# 设置任务专属参数
init_params = {
'basic_shape': task_sticher.basic_shape,
'cut_shape': task_sticher.cut_shape,
'x_area': task_sticher.x_area,
'save_path': task_sticher.save_path,
'train_num': task_sticher.train_num_base + len(stich_tasks),
'type': task_sticher.type
}
# 延迟初始化
if hasattr(task_sticher, 'initialize'):
task_sticher.initialize(init_params)
# 提交任务
future = thread_pool.submit(
stich_worker_thread,
task_imgs,
task_dets,
task_cuts,
task_sticher
)
stich_futures.append(future)
stich_tasks.append({
'end_index': end,
'train_num': self.train_num_base + len(stich_tasks)
})
# 更新数据结构
img_list = img_list[end:]
det_res_list = det_res_list[end:]
img_num_list = img_num_list[end:]
cut = [[0, cut[-1][1]]]
cg_list = [[], []]
# 等待所有剩余任务完成
for future in stich_futures:
try:
res = future.result(timeout=300) # 5分钟超时
if res is not None:
total_res.append(res)
except TimeoutError:
logging.warning(f"{self.task_id} 拼接任务超时")
except Exception as e:
logging.error(f"{self.task_id} 拼接任务失败: {str(e)}")
traceback.print_exc()
# 关闭线程池
thread_pool.shutdown(wait=True)
if cap is not None:
cap.release()
logging.info(f"{self.task_id} 释放视频捕获资源")
return total_res
except Exception as e:
logging.error(f"{self.task_id} 运行异常: {str(e)}")
traceback.print_exc()
# 尝试关闭线程池
try:
thread_pool.shutdown(wait=False, cancel_futures=True)
except:
pass
return total_res上面的代码中，主要通过：
提交任务到线程池
future = thread_pool.submit(
stich_worker_thread,
task_imgs,
task_dets,
task_cuts,
task_sticher
)
来调用拼接程序。
3、基于get_img_asyncio.py程序中给出一个全局函数stich_worker_thread
python
def stich_worker_thread(imgs, dets, cuts, sticher):
try:
return sticher.stich_images(imgs, dets, cuts)
except Exception as e:
logging.error(f"拼接线程出错: {str(e)}")
traceback.print_exc()
return None4、在TotalStich_side和TotalStich_up类中添加initialize方法
python
def initialize(self, init_params):
"""初始化线程专有状态"""
for key, value in init_params.items():
setattr(self, key, value)
# 初始化需要延迟创建的资源
self.weight_map = self.create_weight_map()
self.img_list = []
self.cut = False
# 其他需要重置的状态..