.png%3Ftable%3Dblock%26id%3D99172ae2-ee44-4671-9354-b020520c82ba%26cache%3Dv2&w=3840&q=75)
.png%3Ftable%3Dblock%26id%3Df782c3b9-e714-4386-801d-0fa317203f55%26cache%3Dv2&w=3840&q=75)
0. 准备虚拟环境安装对应的库按键标签准备class分类训练1. 屏幕抓取2. 主程序报错ImportError: DLL load failed while importing win32ui: 找不到指定的程序。解决屏幕识别DNF使用
0. 准备
虚拟环境
启动
# 1.切换到虚拟环境目录 PS D:\Cumtb_Code> cd .\yolo5env\ PS D:\Cumtb_Code\yolo5env> cd .\yolov5-master\ # 2. CMD PS D:\Cumtb_Code\yolo5env\yolov5-master> cmd # 3. 启动虚拟环境 D:\Cumtb_Code\yolo5env\yolov5-master>conda activate D:\Cumtb_Code\yolo5env
安装对应的库
- pywin32（建议安装303版本，304有点小问题，具体是ImportError: DLL load failed while importing win32ui: 找不到指定的程序。）
- PyAutoGUI
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple pywin32==303 pip install -i https://pypi.tuna.tsinghua.edu.cn/simple PyAutoGUI pip install -i https://pypi.tuna.tsinghua.edu.cn/simple pydirectinput-rgx
按键
- 一开始我完全使用的是pywin32，游戏内不可以使用！dnf反外挂屏蔽虚拟按键！
- 按键可以通过win32和ctype同时协作，引入了ctype做虚拟按键，也失效了
- 功夫不负有心人，找到了使用
pydirectinput-rgx库可以操作按键，但鼠标时间在游戏内仍然无法使用
标签准备
labelImg
class分类
0 qigong 1 door 2 monster1 3 monster2 4 monster3 5 monster4 6 boss 7 material 8 money 9 box 10 option 11 ban 12 reddoor 13 jianying 14 renying
训练
python train.py --data ./data/dnf.yaml --cfg ./models/yolov5s.yaml --weights '' --epoch 100 --device 0 python train.py --data ./data/dnf.yaml --cfg ./models/yolov5m.yaml --weights '' --epoch 300 --device 0 python train.py --data ./data/dnf.yaml --cfg ./models/yolov5m.yaml --weights ./runs/train/exp6/weights/best.pt --epoch 300 --device 0
1. 屏幕抓取
使用pywin32对屏幕进行实时抓取
grabscreen.py
import cv2 import numpy as np import win32gui import win32ui import win32con import win32api def grab_screen(region=None): hwin = win32gui.GetDesktopWindow() if region: left, top, x2, y2 = region width = x2 - left + 1 height = y2 - top + 1 else: width = win32api.GetSystemMetrics(win32con.SM_CXVIRTUALSCREEN) height = win32api.GetSystemMetrics(win32con.SM_CYVIRTUALSCREEN) left = win32api.GetSystemMetrics(win32con.SM_XVIRTUALSCREEN) top = win32api.GetSystemMetrics(win32con.SM_YVIRTUALSCREEN) hwindc = win32gui.GetWindowDC(hwin) srcdc = win32ui.CreateDCFromHandle(hwindc) memdc = srcdc.CreateCompatibleDC() bmp = win32ui.CreateBitmap() bmp.CreateCompatibleBitmap(srcdc, width, height) memdc.SelectObject(bmp) memdc.BitBlt((0, 0), (width, height), srcdc, (left, top), win32con.SRCCOPY) signedIntsArray = bmp.GetBitmapBits(True) # img = np.fromstring(signedIntsArray, dtype='uint8') img = np.frombuffer(signedIntsArray, dtype='uint8') img.shape = (height, width, 4) srcdc.DeleteDC() memdc.DeleteDC() win32gui.ReleaseDC(hwin, hwindc) win32gui.DeleteObject(bmp.GetHandle()) return cv2.cvtColor(img, cv2.COLOR_BGRA2RGB)
2. 主程序
- gamemain.py
import random import mss import pydirectinput import win32api,win32con from models.experimental import attempt_load import torch import numpy as np from utils.general import non_max_suppression, scale_coords from utils.augmentations import letterbox import argparse import time import cv2 # import pyautogui parser = argparse.ArgumentParser() # parser.add_argument('--model-path', type=str, default=r'D:\Cumtb_Code\yolo5env\yolov5-master\pretrained\yolov5s.pt', # help='模型地址,绝对路径') parser.add_argument('--model-path', type=str, default='runs/train/exp4/weights/best.pt', help='模型地址,绝对路径') # parser.add_argument('--model-path', type=str, default='pretrained/yolov5s.pt', # help='模型地址,绝对路径') parser.add_argument('--imgsz', type=int, default=640, help='和你训练模型时imgsz一样,默认640') parser.add_argument('--conf-thres', type=float, default=0.1, help='置信阈值') parser.add_argument('--iou-thres', type=float, default=0.05, help='交并比阈值') parser.add_argument('--hide_labels', type=bool, default=True, help='是否隐藏标签') parser.add_argument('--hide_conf', type=bool, default=True, help='是否隐藏置信度') parser.add_argument('--show-window', type=bool, default=True, help='是否显示实时检测窗口') parser.add_argument('--resize-window', type=float, default=1 / 4, help='缩放实时检测窗口大小') parser.add_argument('--show-fps', type=bool, default=True, help='是否显示帧数') parser.add_argument('--region', type=tuple, default=(1, 1), help='检测范围；分别为x，y，(1.0, 1.0)表示全屏检测，越低检测范围越小(始终保持屏幕中心为中心)') args = parser.parse_args() '------------------------------------------------------------------------------------' # 加载模型 def load_model(args): device = 'cuda' if torch.cuda.is_available() else 'cpu' # 设备选择 half = device != 'cpu' # fp32/fp16 model = attempt_load(args.model_path, device=device) # 加载 FP32 模型 if half: # 如果cuda可用 model.half() # 启用 FP16 if device != 'cpu': # 如果是cuda model(torch.zeros(1, 3, args.imgsz, args.imgsz).to(device).type_as(next(model.parameters()))) # cuda设置 return model # 返回加载好的网络模型 # mss截图 cap = mss.mss() # 实例化mss def grab_screen_mss(monitor): # cap.grab截取图片，np.array将图片转为数组，cvtColor将BRGA转为BRG,去掉了透明通道 return cv2.cvtColor(np.array(cap.grab(monitor)), cv2.COLOR_BGRA2BGR) # game skill_char = "ASDFGHQWERT" # 技能按键，使用均匀分布随机抽取 # material_xywh = ['0', '0.768555', '910', '500', '1110', '580'] hero_buff = 0 monster_show = [0,0,0,0,0] hero_xywh_cache = [0, 450, 110, 450] role_time = 0 no_time = 0 ban_time = 0 # 运行 def run(): # top_x, top_y, x, y = 0, 0, 1920, 1080 # x,y 屏幕大小,top是原点 top_x, top_y, x, y = 0, 0, 1280, 720 # x,y 屏幕大小,top是原点 len_x, len_y = int(x * args.region[0]), int(y * args.region[1]) # 截图的宽高 top_x, top_y = int(top_x + x // 2 * (1. - args.region[0])), int(top_y + y // 2 * (1. - args.region[1])) # 截图区域的原点 monitor = {'left': top_x, 'top': top_y, 'width': len_x, 'height': len_y} # 截图范围 cv2.namedWindow('img', cv2.WINDOW_NORMAL) # 创建窗口 if args.show_window: # 是否显示检测款 len_x, len_y = int(x * args.region[0]), int(y * args.region[1]) cv2.resizeWindow('img', int(len_x * args.resize_window), int(len_y * args.resize_window)) # 裁剪窗口 t0 = time.time() # fps 计算 while True: if not cv2.getWindowProperty('img', cv2.WND_PROP_VISIBLE): # 如果窗口关闭，退出程序 cv2.destroyAllWindows() exit('程序结束...') break img0 = grab_screen_mss(monitor) # 截取整个屏幕的到图片img0 img0 = cv2.resize(img0, (len_x, len_y)) # 裁剪图片至截取的大小 # 预处理 img = letterbox(img0, args.imgsz, stride=stride)[0] # 预处理 img = img.transpose((2, 0, 1))[::-1] # 维度转换 img = np.ascontiguousarray(img) # 转为数组，其内存是连续的 img = torch.from_numpy(img).to(device) # 将来自numpy的数组转为tensor，并传入设备 img = img.half() if half else img.float() # 选择fp32 / fp16 img /= 255. # 归一化 ,0 - 255 to 0.0 - 1.0 img = img[None] # 扩大批调暗 # 推理 t1 = time.time() # 时间点 pred = model(img, augment=False, visualize=False)[0] # Apply NMS,非极大值抑制 pred = non_max_suppression(pred, conf_thres, iou_thres, agnostic=False) t2 = time.time() # print('推理时间 {} ms'.format('%.2f' % ((t2 - t1) * 1000))) # Process detections,转换 aims = [] for i, det in enumerate(pred): # detections per image s, im0 = '', img0.copy() # 输出字符串 s += '%gx%g ' % img.shape[2:] global no_time if len(det): # Rescale boxes from img_size to im0 size,将坐标 (xyxy) 从 img_shape 重新缩放为 img0_shape det[:, :4] = scale_coords(img.shape[2:], det[:, :4], img0.shape).round() # Write results for *xyxy, conf, cls in reversed(det): # 从末尾遍历 # 将xyxy合并至一个维度,锚框的左上角和右下角 xyxy = (torch.tensor(xyxy).view(1, 4)).view(-1) # 将类别和坐标合并 line = (cls,conf, *xyxy) # 提取tensor类型里的坐标数据 aim = ('%g ' * len( line)).rstrip() % line # %g 格式为浮点数 .rstrip()删除tring字符串末尾的指定字符,默认为空白符包括空格,即删除2个坐标之间的空格 # 划分元素 aim = aim.split(' ') # 将一个元素按空格符分为多个元素,获得单个目标信息列表 # 所有目标的类别和锚框的坐标(类别,左上角x,左上角y,右下角x,右下角y) aims.append(aim) # 添加至列表 # aims.append(aim) # 加入标签列表 if len(aims): # 如果检测到存在目标 font = cv2.FONT_HERSHEY_SIMPLEX cls_object = [] # 名称列表 loc_list = [] # 坐标 global hero_buff global monster_show global role_time global ban_time global hero_xywh_cache # hero_conf = 0 hero_index = 0 for idx, det in enumerate(aims): cls, conf,x_center, y_center, width, height = det # 将det里的数据分装到前面 rc_x,y 表示归化后的比例坐标 top_left = (int(x_center), int(y_center)) bottom_right = (int(width), int(height)) cls = int(cls) conf = float(conf) cv2.rectangle(img0, top_left, bottom_right, color=colors[cls], thickness=3) # 3代表线条粗细 cv2.putText(img0, names[cls], (top_left[0], top_left[1] - 5), font, 1, colors[cls], 2) cls_object.append(names[cls]) xywh = [int(x_center), int(y_center),int(width), int(height)] loc_list.append(xywh) # print(cls_object) if names[cls] == "role": # hero_conf = conf hero_index = idx # 加上BUFF if hero_buff <= 0: print("加上BuFF") pydirectinput.keyDown('ctrl') time.sleep(0.02) pydirectinput.keyUp('ctrl') time.sleep(2) pydirectinput.keyDown('alt') time.sleep(0.04) pydirectinput.keyUp('alt') time.sleep(0.01) # 进入下一个房间 pydirectinput.keyDown('right') time.sleep(0.02) pydirectinput.keyUp('right') time.sleep(0.01) pydirectinput.keyDown('right') time.sleep(2) pydirectinput.keyUp('right') time.sleep(0.01) hero_buff +=1 # 扫描英雄 if "role" in cls_object: # hero_xywh = aims[hero_index] hero_xywh = loc_list[hero_index] hero_xywh_cache = hero_xywh else: # continue hero_xywh = hero_xywh_cache # print('没有扫到角色') # 打怪 # 阈值设置 thx = 30 # 捡东西时，x方向的阈值 thy = 30 # 捡东西时，y方向的阈值 attx=[400,520,420,300,550] # 攻击时，x方向的阈值 atty=[180,300,150,100,500] att_key = np.random.randint(len(skill_char), size=2) # skill_list = ["A","S","D","F","G","H","Q","W","E","R","T","Y"] monster=["monster1",'monster2','monster3','monster4','boss'] # 判断是否出现目标，打怪，我这里设置的是两个技能秒怪！！！！ for j,mon in enumerate(monster): if mon in cls_object: index = cls_object.index(mon) monster_xywh = loc_list[index] # 怪物和英雄的距离绝对值,释放技能 if monster_show[j] <= 0: # print("默认像怪物右移动一段距离") # hero_xywh = [0, 450, 110, 450] # hero_xywh[1] = monster_xywh[1] # hero_xywh[3] = monster_xywh[3] # pydirectinput.keyDown('down') # time.sleep(0.2) # pydirectinput.keyUp('down') # time.sleep(0.01) pydirectinput.keyDown('right') time.sleep(0.4) pydirectinput.keyUp('right') time.sleep(0.01) monster_show[j] += 1 if abs((monster_xywh[2]+ monster_xywh[0])/2 - (hero_xywh[2]+ hero_xywh[0])/2) <= attx[j] and -atty[j]-100<= (monster_xywh[3]+monster_xywh[1])/2 - (hero_xywh[3]+hero_xywh[1])/2 <= atty[j]: if (monster_xywh[2]+monster_xywh[0])/2 - (hero_xywh[2]+hero_xywh[0])/2 > 0: pydirectinput.keyDown('right') time.sleep(0.02) pydirectinput.keyUp('right') time.sleep(0.01) print("怪物在右边") if mon == "boss": skill_name1 = "Y" else: skill_name1 = skill_char[int(att_key[0])] # skill_name1 = skill_char[int(att_key[0])] pydirectinput.keyDown(skill_name1) # 按下键盘 time.sleep(0.2) pydirectinput.keyUp(skill_name1) # 松开键盘 time.sleep(0.01) else: pydirectinput.keyDown('left') time.sleep(0.02) pydirectinput.keyUp('left') time.sleep(0.01) print("怪物在左边") if mon == "boss": skill_name1 = "Y" else: skill_name1 = skill_char[int(att_key[0])] skill_name1 = skill_char[int(att_key[0])] pydirectinput.keyDown(skill_name1) # 按下键盘 time.sleep(0.2) pydirectinput.keyUp(skill_name1) # 松开键盘 time.sleep(0.01) else: if (monster_xywh[2]+ monster_xywh[0])/2 - (hero_xywh[2]+ hero_xywh[0])/2 > attx[j] : pydirectinput.keyDown('right') time.sleep(0.04) pydirectinput.keyUp('right') time.sleep(0.01) print(mon,"向右靠近怪物") else: pydirectinput.keyDown('left') time.sleep(0.04) pydirectinput.keyUp('left') time.sleep(0.01) print(mon,"向左靠近怪物") if (monster_xywh[3]+monster_xywh[1])/2 - (hero_xywh[3] + hero_xywh[1])/2> atty[j]: pydirectinput.keyDown('down') time.sleep(0.04) pydirectinput.keyUp('down') time.sleep(0.01) print(mon,"向下靠近怪物") else: pydirectinput.keyDown('up') time.sleep(0.04) pydirectinput.keyUp('up') time.sleep(0.01) print(mon,"向上靠近怪物") # 有异形空间 if "ban" in cls_object: if ban_time <=1: pydirectinput.keyDown('down') time.sleep(1.5) pydirectinput.keyUp('down') time.sleep(0.01) ban_time +=1 else: pydirectinput.keyDown('right') time.sleep(4) pydirectinput.keyUp('right') time.sleep(0.01) # 捡材料 if "monster1" not in cls_object and "monster2" not in cls_object and "monster3" not in cls_object and "monster4" not in cls_object and "boss" not in cls_object and ("material" in cls_object or "money" in cls_object) and "ban" not in cls_object: material_xywh_list = [] if "material" in cls_object: index = cls_object.index("material") material_xywh_list.append(loc_list[index]) elif "money" in cls_object: index = cls_object.index("money") material_xywh_list.append(loc_list[index]) else: print("继续扫描物品") for material_xywh in material_xywh_list: # print("material_xywh",material_xywh) if abs((material_xywh[2] + material_xywh[0])/2 - hero_xywh[2]) <= thx and abs((material_xywh[3] + material_xywh[1])/2 - hero_xywh[3])<= thy: # 按X捡材料 pydirectinput.keyDown('X') time.sleep(0.02) pydirectinput.keyUp('X') time.sleep(0.01) print("可以捡到东西") else: if (material_xywh[2] + material_xywh[0])/2 - hero_xywh[2] > thx: pydirectinput.keyDown('right') time.sleep(0.24) pydirectinput.keyUp('right') time.sleep(0.01) print("向右移动可以捡到东西") else: pydirectinput.keyDown('left') time.sleep(0.24) pydirectinput.keyUp('left') time.sleep(0.01) print("向左移动可以捡到东西") if (material_xywh[3] + material_xywh[1])/2 - hero_xywh[3] > thy: pydirectinput.keyDown('down') time.sleep(0.24) pydirectinput.keyUp('down') time.sleep(0.01) print("向下移动可以捡到东西") else: pydirectinput.keyDown('up') time.sleep(0.24) pydirectinput.keyUp('up') time.sleep(0.01) print("向上移动可以捡到东西") # 进入下一个房间,有门 if "door" in cls_object and "monster1" not in cls_object and "monster2" not in cls_object and "monster3" not in cls_object and "monster4" not in cls_object and "material" not in cls_object and "money" not in cls_object and "ban" not in cls_object: index = cls_object.index("door") door_xywh = loc_list[index] door_x = (door_xywh[2]+door_xywh[0])/2 if door_x < 900: print("门在左边") pydirectinput.keyDown('right') time.sleep(0.02) pydirectinput.keyUp('right') time.sleep(0.01) pydirectinput.keyDown('right') time.sleep(1.5) pydirectinput.keyUp('right') time.sleep(0.01) else: print("门在右边") if (door_xywh[3]+door_xywh[1])/2 - (hero_xywh[3]+hero_xywh[1])/2 >= 40: pydirectinput.keyDown('up') time.sleep(0.04) pydirectinput.keyUp('up') time.sleep(0.01) pydirectinput.keyDown('down') time.sleep(0.08) pydirectinput.keyUp('down') time.sleep(0.01) print("出现门，向下移动") elif (door_xywh[3]+door_xywh[1])/2 - (hero_xywh[3]+hero_xywh[1])/2 <= -10: pydirectinput.keyDown('down') time.sleep(0.04) pydirectinput.keyUp('down') time.sleep(0.01) pydirectinput.keyDown('up') time.sleep(0.08) pydirectinput.keyUp('up') time.sleep(0.01) print("出现门，向上移动") else: print("出现门，向右移动") pydirectinput.keyDown('left') time.sleep(0.4) pydirectinput.keyUp('left') time.sleep(0.01) pydirectinput.keyDown('right') time.sleep(0.02) pydirectinput.keyUp('right') time.sleep(0.01) pydirectinput.keyDown('right') time.sleep(0.8) pydirectinput.keyUp('right') time.sleep(0.01) # 只有角色在 if "role" in cls_object and "door" not in cls_object and "monster1" not in cls_object and "monster2" not in cls_object and "monster3" not in cls_object and "monster4" not in cls_object and "boss" not in cls_object and "material" not in cls_object and "money" not in cls_object and "option" not in cls_object and "box" not in cls_object and "ban" not in cls_object: print("只有角色在") if role_time == 0: role_time = time.time() if time.time()- role_time >=5: pydirectinput.keyDown('left') time.sleep(0.4) pydirectinput.keyUp('left') time.sleep(0.01) role_time = 0 # 开箱子 if "box" in cls_object: box_num = 0 for b in cls_object: if b == "box": box_num += 1 if box_num >= 4: pydirectinput.keyDown('esc') time.sleep(0.02) pydirectinput.keyUp('esc') time.sleep(0.1) print("打开箱子ESC") pydirectinput.keyDown('f3') time.sleep(0.02) pydirectinput.keyUp('f3') time.sleep(0.01) # 重新开始 time_option = -20 if "option" in cls_object: if time.time() - time_option > 10: # pydirectinput.moveTo(540,450) pydirectinput.keyDown('esc') time.sleep(0.02) pydirectinput.keyUp('esc') time.sleep(0.1) pydirectinput.keyDown('f12') time.sleep(0.02) pydirectinput.keyUp('f12') time.sleep(1) pydirectinput.keyDown('left') time.sleep(0.2) pydirectinput.keyUp('left') time.sleep(1) pydirectinput.keyDown('right') time.sleep(2) pydirectinput.keyUp('right') time.sleep(0.01) pydirectinput.keyDown('space') time.sleep(0.02) pydirectinput.keyUp('space') time.sleep(1) monster_show = [0,0,0,0,0] hero_buff = 0 ban_time = 0 time_option = time.time() # 什么都没检测到 else: print("什么都没检测到") if no_time == 0: no_time = time.time() if time.time()- no_time >=5: pydirectinput.keyDown('left') time.sleep(0.4) pydirectinput.keyUp('left') time.sleep(0.01) no_time = 0 # 显示检测 if args.show_window: # 是否显示窗口 if args.show_fps: # 是否显示 fps cv2.putText(img0, "FPS:{:.1f}".format(1. / (time.time() - t0)), (50, 50), cv2.FONT_HERSHEY_SIMPLEX, 2, (0, 0, 235), 4) # 绘制字体 t0 = time.time() cv2.imshow('img', img0) # 显示 cv2.waitKey(1) if __name__ == '__main__': # 参数初始化 model = load_model(args) # 加载模型 stride = int(model.stride.max()) # 设置特征点步长 # 获取类名 names = model.module.names if hasattr(model, 'module') else model.names # 设置边框颜色 colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(names))] device = 'cuda' if torch.cuda.is_available() else 'cpu' # 根据pytorch选择设备，cpu或者cuda conf_thres = args.conf_thres # 置信度 iou_thres = args.iou_thres # IOU half = device != 'cpu' # 如果cuda可用，启用fp16 # run run()
报错
ImportError: DLL load failed while importing win32ui: 找不到指定的程序。
解决
我安装的是pywin32的304版本，卸载，再装303版本即可
pip install pywin32 pip install -i https://pypi.tuna.tsinghua.edu.cn/simple pywin32==303
屏幕识别
DNF使用
我们将训练好的yolov5模型加载，然后通过实时的屏幕抓取对每一帧图像做目标检测，如果检测到图片中有怪物和自身角色，那么记录自己的坐标和怪物的坐标，通过虚拟按键向怪物所在位置移动；如果怪物没清空了并且地上有掉落的材料，那么控制角色靠近材料所在的坐标；如果怪物和材料都不存在，那么控制角色走向传送门的位置，即走向下一个房间，具体细节和原理在b站视频中有详细的介绍，这里给出部分代码：
import mss from models.experimental import attempt_load import torch import numpy as np from utils.general import non_max_suppression, scale_coords, xyxy2xywh from utils.augmentations import letterbox import argparse import time import cv2 # opencv-python 不要超过4.5 ''' v6.0版本训练出来的模型必须和同版本神经网络一致，否则会报错 ''' parser = argparse.ArgumentParser() parser.add_argument('--model-path', type=str, default=r'C:\Users\Zzzz\Desktop\yolov5-6.0\weights\cf_v6.pt', help='模型地址,绝对路径') parser.add_argument('--imgsz', type=int, default=640, help='和你训练模型时imgsz一样,默认640') parser.add_argument('--conf-thres', type=float, default=0.1, help='置信阈值') parser.add_argument('--iou-thres', type=float, default=0.05, help='交并比阈值') parser.add_argument('--show-window', type=bool, default=True, help='是否显示实时检测窗口') parser.add_argument('--resize-window', type=float, default=1 / 2, help='缩放实时检测窗口大小') parser.add_argument('--show-fps', type=bool, default=True, help='是否显示帧数') parser.add_argument('--region', type=tuple, default=(0.5, 0.5), help='检测范围；分别为x，y，(1.0, 1.0)表示全屏检测，越低检测范围越小(始终保持屏幕中心为中心)') args = parser.parse_args() '------------------------------------------------------------------------------------' # 加载模型 def load_model(args): device = 'cuda' if torch.cuda.is_available() else 'cpu' # 设备选择 half = device != 'cpu' # fp32/fp16 model = attempt_load(args.model_path, map_location=device) # 加载 FP32 模型 if half: # 如果cuda可用 model.half() # 启用 FP16 if device != 'cpu': # 如果是cuda model(torch.zeros(1, 3, args.imgsz, args.imgsz).to(device).type_as(next(model.parameters()))) # cuda设置 return model # 返回加载好的网络模型 # mss截图 cap = mss.mss() # 实例化mss def grab_screen_mss(monitor): # cap.grab截取图片，np.array将图片转为数组，cvtColor将BRGA转为BRG,去掉了透明通道 return cv2.cvtColor(np.array(cap.grab(monitor)), cv2.COLOR_BGRA2BGR) # 画框函数 def fun_en(aims, img0, len_x, len_y): for i, det in enumerate(aims): _, x_center, y_center, width, height = det # 将det里的数据分装到前面 rc_x,y 表示归化后的比例坐标 x_center, width = len_x * float(x_center), len_x * float(width) # 中心的x和宽 y_center, height = len_y * float(y_center), len_y * float(height) # 中心的y和高 top_left = (int(x_center - width / 2.), int(y_center - height / 2.)) bottom_right = (int(x_center + width / 2.), int(y_center + height / 2.)) color = (0, 255, 0) # RGB 框的颜色 cv2.rectangle(img0, top_left, bottom_right, color, thickness=3) # 3代表线条粗细 # 运行 def run(): top_x, top_y, x, y = 0, 0, 1920, 1080 # x,y 屏幕大小,top是原点 len_x, len_y = int(x * args.region[0]), int(y * args.region[1]) # 截图的宽高 top_x, top_y = int(top_x + x // 2 * (1. - args.region[0])), int(top_y + y // 2 * (1. - args.region[1])) # 截图区域的原点 monitor = {'left': top_x, 'top': top_y, 'width': len_x, 'height': len_y} # 截图范围 cv2.namedWindow('img', cv2.WINDOW_NORMAL) # 创建窗口 if args.show_window: # 是否显示检测款 len_x, len_y = int(x * args.region[0]), int(y * args.region[1]) cv2.resizeWindow('img', int(len_x * args.resize_window), int(len_y * args.resize_window)) # 裁剪窗口 t0 = time.time() # fps 计算 while True: if not cv2.getWindowProperty('img', cv2.WND_PROP_VISIBLE): # 如果窗口关闭，退出程序 cv2.destroyAllWindows() exit('程序结束...') break img0 = grab_screen_mss(monitor) # 截取整个屏幕的到图片img0 img0 = cv2.resize(img0, (len_x, len_y)) # 裁剪图片至截取的大小 # 预处理 img = letterbox(img0, args.imgsz, stride=stride)[0] # 预处理 img = img.transpose((2, 0, 1))[::-1] # 维度转换 img = np.ascontiguousarray(img) # 转为数组，其内存是连续的 img = torch.from_numpy(img).to(device) # 将来自numpy的数组转为tensor，并传入设备 img = img.half() if half else img.float() # 选择fp32 / fp16 img /= 255. # 归一化 img = img[None] # 扩大批调暗 # if len(img.shape) == 3: # img = img[None] # 推理 t1 = time.time() # 时间点 pred = model(img, augment=False, visualize=False)[0] # 非极大值抑制 pred = non_max_suppression(pred, conf_thres, iou_thres, agnostic=False) t2 = time.time() print('推理时间 {} ms'.format('%.2f' % ((t2 - t1) * 1000))) # 转换 aims = [] for i, det in enumerate(pred): if len(det): # 将坐标 (xyxy) 从 img_shape 重新缩放为 img0_shape det[:, :4] = scale_coords(img.shape[2:], det[:, :4], img0.shape).round() for *xyxy, conf, cls in reversed(det): # 从末尾遍历 # 将xyxy合并至一个维度,锚框的左上角和右下角 xyxy = (torch.tensor(xyxy).view(1, 4)).view(-1) # 将类别和坐标合并 line = (cls, *xyxy) # 提取tensor类型里的坐标数据 aim = ('%g ' * len( line)).rstrip() % line # %g 格式为浮点数 .rstrip()删除tring字符串末尾的指定字符,默认为空白符包括空格,即删除2个坐标之间的空格 # 划分元素 aim = aim.split(' ') # 将一个元素按空格符分为多个元素,获得单个目标信息列表 # 所有目标的类别和锚框的坐标(类别,左上角x,左上角y,右下角x,右下角y) aims.append(aim) # 添加至列表 aims.append(aim) # 加入标签列表 if len(aims): # 如果检测到存在目标 fun_en(aims, img0, len_x, len_y) # 画框函数 # 显示检测 if args.show_window: # 是否显示窗口 if args.show_fps: # 是否显示 fps cv2.putText(img0, "FPS:{:.1f}".format(1. / (time.time() - t0)), (50, 50), cv2.FONT_HERSHEY_SIMPLEX, 2, (0, 0, 235), 4) # 绘制字体 t0 = time.time() cv2.imshow('img', img0) # 显示 cv2.waitKey(1) if __name__ == '__main__': # 参数初始化 model = load_model(args) # 加载模型 stride = int(model.stride.max()) # 设置特征点步长 device = 'cuda' if torch.cuda.is_available() else 'cpu' # 根据pytorch选择设备，cpu或者cuda conf_thres = args.conf_thres # 置信度 iou_thres = args.iou_thres # IOU half = device != 'cpu' # 如果cuda可用，启用fp16 # run run()