修改YOLOv5 detect.py代码使其能逐个视频检测保存，同时对每个视频内参数进行单独操作

真没怎么看懂YOLOv5的detect.py代码的逻辑，看了YOLOv3，和YOLOv4的detect逻辑，基本都是用opencv对每个视频进行操作，感觉还清晰易懂一点，YOLOv5的作者都好像没用opencv进行操作，或者把opencv的视频操作封装成另一个py文件隐藏起来，实在有些隐晦，所以用了最笨的方法，用os.listdir读视频文件目录下的所有视频，逐一检测。同时改写了画框的函数(因为要保存一帧关键帧的内容)，检测命令里是用python detect.py --exist-ok --nosave，因为检测命令里带nosave这一选项，所以浅扒了一下作者的画框逻辑，发现还是用的opencv的rectangle方法(作者藏的

import numpy as np
import argparse
import os
import sys
from pathlib import Path
import time
import shutil
from PIL import Image
import cv2
import torch
import torch.backends.cudnn as cudnn

FILE = Path(__file__).resolve()
ROOT = FILE.parents[0]  # YOLOv5 root directory
if str(ROOT) not in sys.path:
    sys.path.append(str(ROOT))  # add ROOT to PATH
ROOT = Path(os.path.relpath(ROOT, Path.cwd()))  # relative

from models.common import DetectMultiBackend
from utils.datasets import IMG_FORMATS, VID_FORMATS, LoadImages, LoadStreams
from utils.general import (LOGGER, check_file, check_img_size, check_imshow, check_requirements, colorstr,
                           increment_path, non_max_suppression, print_args, scale_coords, strip_optimizer, xyxy2xywh)
from utils.plots import Annotator, colors, save_one_box
from utils.torch_utils import select_device, time_sync

@torch.no_grad()
def run(weights=ROOT / 'yolov5s.pt',  # model.pt path(s)
        vidpath='/home/ccf_disk/animal/test/',  # file/dir/URL/glob, 0 for webcam
        data=ROOT / 'data/coco128.yaml',  # dataset.yaml path
        imgsz=(640, 640),  # inference size (height, width)
        conf_thres=0.6,  # confidence threshold
        iou_thres=0.45,  # NMS IOU threshold
        max_det=1000,  # maximum detections per image
        device='',  # cuda device, i.e. 0 or 0,1,2,3 or cpu
        view_img=False,  # show results
        save_txt=False,  # save results to *.txt
        save_conf=False,  # save confidences in --save-txt labels
        save_crop=False,  # save cropped prediction boxes
        nosave=True,  # do not save images/videos
        classes=None,  # filter by class: --class 0, or --class 0 2 3
        agnostic_nms=False,  # class-agnostic NMS
        augment=False,  # augmented inference
        visualize=False,  # visualize features
        update=False,  # update all models
        project='/home/ccf_disk/animal/video_animal',  # save results to project/name
        name='test_1',  # save results to project/name
        exist_ok=True,  # existing project/name ok, do not increment
        line_thickness=3,  # bounding box thickness (pixels)
        hide_labels=False,  # hide labels
        hide_conf=False,  # hide confidences
        half=False,  # use FP16 half-precision inference
        dnn=False,  # use OpenCV DNN for ONNX inference
        ):
    vidpath = str(vidpath)
    videos = os.listdir(vidpath)
    number = 0
    for video_name in videos:
        time1_start = time.time()
        so = vidpath + video_name
        number = number + 1
        print("第%d个视频处理中" %number)
        source = str(so)
        save_c = 0
        keep = 0
        save_img = not nosave and not source.endswith('.txt')  # save inference images
        is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS)
        is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
        webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file)
        if is_url and is_file:
            source = check_file(source)  # download

        # Directories
        save_dir = increment_path(Path(project) / name, exist_ok=exist_ok)  # increment run
        (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make dir

        # Load model
        device = select_device(device)
        model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data)
        stride, names, pt, jit, onnx, engine = model.stride, model.names, model.pt, model.jit, model.onnx, model.engine
        imgsz = check_img_size(imgsz, s=stride)  # check image size

        # Half
        half &= (pt or jit or onnx or engine) and device.type != 'cpu'  # FP16 supported on limited backends with CUDA
        if pt or jit:
            model.model.half() if half else model.model.float()

        # Dataloader
        if webcam:
            view_img = check_imshow()
            cudnn.benchmark = True  # set True to speed up constant image size inference
            dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt)
            bs = len(dataset)  # batch_size
        else:
            dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt)
            bs = 1  # batch_size
        vid_path, vid_writer = [None] * bs, [None] * bs

        # Run inference
        model.warmup(imgsz=(1 if pt else bs, 3, *imgsz), half=half)  # warmup
        dt, seen = [0.0, 0.0, 0.0], 0
        for path, im, im0s, vid_cap, s in dataset:
            flag = 0
            c = 1
            time1 = 6
            # t1 = time_sync()
            im = torch.from_numpy(im).to(device)
            im = im.half() if half else im.float()  # uint8 to fp16/32
            im /= 255  # 0 - 255 to 0.0 - 1.0
            if len(im.shape) == 3:
                im = im[None]  # expand for batch dim
            # t2 = time_sync()
            # dt[0] += t2 - t1

            # Inference

            visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False
            pred = model(im, augment=augment, visualize=visualize)
            # t3 = time_sync()
            # dt[1] += t3 - t2

            # NMS
            pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det)

            # dt[2] += time_sync() - t3

            # Second-stage classifier (optional)
            # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s)

            # Process predictions
            for i, det in enumerate(pred):  # per image
                seen += 1
                count = 0
                if webcam:  # batch_size >= 1
                    p, im0, frame = path[i], im0s[i].copy(), dataset.count
                    s += f'{i}: '
                else:
                    p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0)

                p = Path(p)  # to Path
                save_path = str(save_dir / p.name)  # im.jpg
                txt_path = str(save_dir / 'labels' / p.stem) + (
                    '' if dataset.mode == 'image' else f'_{frame}')  # im.txt
                s += '%gx%g ' % im.shape[2:]  # print string
                gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  # normalization gain whwh
                imc = im0.copy() if save_crop else im0  # for save_crop
                annotator = Annotator(im0, line_width=line_thickness, example=str(names))
                if len(det):
                    # Rescale boxes from img_size to im0 size
                    det[:, :4] = scale_coords(im.shape[2:], det[:, :4], im0.shape).round()

                    # Print results
                    for c in det[:, -1].unique():
                        n = (det[:, -1] == c).sum()  # detections per class
                        s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "  # add to string

                    # Write results
                    for *xyxy, conf, cls in reversed(det):
                        count = 1
                        if save_txt:  # Write to file
                            xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywh
                            line = (cls, *xywh, conf) if save_conf else (cls, *xywh)  # label format
                            with open(txt_path + '.txt', 'a') as f:
                                f.write(('%g ' * len(line)).rstrip() % line + '\n')

                        if save_img or save_crop or view_img:  # Add bbox to image
                            c = int(cls)  # integer class
                            label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}')
                            annotator.box_label(xyxy, label, color=colors(c, True))
                            if save_crop:
                                save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True)
                        box = xyxy        
                        c = int(cls)  # integer class
                        p1, p2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3]))
                        lw = max(round(sum(im0.shape) / 2 * 0.003), 2)
                        cv2.rectangle(im0, p1, p2, color=(0, 0, 255),
                                      thickness=max(round(sum(im0.shape) / 2 * 0.003), 2), lineType=cv2.LINE_AA)
                        label = (f'{names[c]} {conf:.2f}')
                        tf = max(lw - 1, 1)
                        w, h = cv2.getTextSize(label, 0, fontScale=lw / 3, thickness=tf)[0]  # text width, height
                        outside = p1[1] - h - 3 >= 0  # label fits outside box
                        cv2.putText(im0, label, (p1[0], p1[1] - 2 if outside else p1[1] + h + 2), 0, lw / 3,
                                    (0, 0, 255),
                                    thickness=tf, lineType=cv2.LINE_AA)

                # Stream results
                im0 = annotator.result()
                if view_img:
                    cv2.imshow(str(p), im0)
                    cv2.waitKey(1)  # 1 millisecond
                if (seen % time1 == 0):
                    if (count == 0):
                        save_c = 0
                    else:
                        save_c = save_c + 1
                if(save_c>=4):
                    if keep == 0:
                        im0 = cv2.cvtColor(im0, cv2.COLOR_BGR2RGB)
                        frame = Image.fromarray(np.uint8(im0))
                        #print(save_path)
                        frame.save(str(save_path.split('.')[0]) + ".jpg")
                        keep = 1
                        shutil.copy(so, save_path)
                        print('have animal')
                        break
            else:
                continue
            break

            # # Save results (image with detections)
            # if save_img:
            #     if dataset.mode == 'image':
            #         cv2.imwrite(save_path, im0)
            #     else:  # 'video' or 'stream'
            #         if vid_path[i] != save_path:  # new video
            #             vid_path[i] = save_path
            #             if isinstance(vid_writer[i], cv2.VideoWriter):
            #                 vid_writer[i].release()  # release previous video writer
            #             if vid_cap:  # video
            #                 fps = vid_cap.get(cv2.CAP_PROP_FPS)
            #                 w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
            #                 h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
            #             else:  # stream
            #                 fps, w, h = 30, im0.shape[1], im0.shape[0]
            #             save_path = str(Path(save_path).with_suffix('.mp4'))  # force *.mp4 suffix on results videos
            #             vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
            #         vid_writer[i].write(im0)

            # Print time (inference-only)
            # LOGGER.info(f'{s}Done. ({t3 - t2:.3f}s)')

        # Print results
        # t = tuple(x / seen * 1E3 for x in dt)  # speeds per image
        # LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t)
        if save_txt or save_img:
            s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
            # LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}")
        if update:
            strip_optimizer(weights)  # update model (to fix SourceChangeWarning)

        time1_end = time.time()
        print('视频%d处理时间' % number + str(time1_end-time1_start))
        # if bool == True:
        #     shutil.copy(so, save_path)
        # else:
        #     pass

def parse_opt():
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'weights/best.pt', help='model path(s)')
    parser.add_argument('--vidpath', type=str, default='/home/ccf_disk/animal/video/4-3/',
                        help='file/dir/URL/glob, 0 for webcam')
    parser.add_argument('--data', type=str, default=ROOT / 'data/myvoc.yaml', help='(optional) dataset.yaml path')
    parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w')
    parser.add_argument('--conf-thres', type=float, default=0.75, help='confidence threshold')
    parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold')
    parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image')
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--view-img', action='store_true', help='show results')
    parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
    parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
    parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes')
    parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
    parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3')
    parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
    parser.add_argument('--augment', action='store_true', help='augmented inference')
    parser.add_argument('--visualize', action='store_true', help='visualize features')
    parser.add_argument('--update', action='store_true', help='update all models')
    parser.add_argument('--project', default='/home/ccf_disk/animal/video_animal_yolov5/', help='save results to project/name')
    parser.add_argument('--name', default='4-3', help='save results to project/name')
    parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
    parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)')
    parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels')
    parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences')
    parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference')
    parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference')
    opt = parser.parse_args()
    opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1  # expand
    print_args(FILE.stem, opt)
    return opt

def main(opt):
    check_requirements(exclude=('tensorboard', 'thop'))
    run(**vars(opt))

if __name__ == "__main__":
    opt = parse_opt()
    main(opt)

有点深)，第一次发博客，浅记录一下。