二次元刀哥镇楼
简介
先用人家代码实现该功能，以后（刀哥保佑我顺利毕业，否则遥遥无期）大概率要：
- 提高检测器的检测能力（数据训练，新的yolo模型）
- 提高deepsort的分辨能力（低视角，网后人员的辨别）
- 利用mmpose提取球员的骨架
- 球员运动轨迹（仿射变换到二维球场上）
- 羽球的跟踪（双目视觉定位）
具体实现过程：
- 使用
query_get.py截取视频中的人物，作为reid的样本 - 加载
yolov3.weight权重 - 调整
search.py部分参数，运行
-
截取图片样例，其中包含了网后的图片，但是在deepsort中仍然无法匹配：
-
其中
search.py的代码：
import argparse
import time
from sys import platform
from models import *
from utils.datasets import *
from utils.utils import *
from reid.data import make_data_loader
from reid.data.transforms import build_transforms
from reid.modeling import build_model
from reid.config import cfg as reidCfg
def detect(cfg,
data,
weights,
images='data/samples', # input folder
output='output', # output folder
fourcc='mp4v', # video codec
img_size=416,
conf_thres=0.5,
nms_thres=0.5,
dist_thres=1.0,
save_txt=False,
save_images=True):
# Initialize
device = torch_utils.select_device(force_cpu=False)
torch.backends.cudnn.benchmark = False # set False for reproducible results
if os.path.exists(output):
shutil.rmtree(output) # delete output folder
os.makedirs(output) # make new output folder
############# 行人重识别模型初始化 #############
query_loader, num_query = make_data_loader(reidCfg)
reidModel = build_model(reidCfg, num_classes=10126)
reidModel.load_param(reidCfg.TEST.WEIGHT)
reidModel.to(device).eval()
query_feats = []
query_pids = []
for i, batch in enumerate(query_loader):
with torch.no_grad():
img, pid, camid = batch
img = img.to(device)
feat = reidModel(img) # 一共2张待查询图片，每张图片特征向量2048 torch.Size([2, 2048])
query_feats.append(feat)
query_pids.extend(np.asarray(pid)) # extend() 函数用于在列表末尾一次性追加另一个序列中的多个值（用新列表扩展原来的列表）。
query_feats = torch.cat(query_feats, dim=0) # torch.Size([2, 2048])
print("The query feature is normalized")
query_feats = torch.nn.functional.normalize(query_feats, dim=1, p=2) # 计算出查询图片的特征向量
############# 行人检测模型初始化 #############
model = Darknet(cfg, img_size)
# Load weights
if weights.endswith('.pt'): # pytorch format
model.load_state_dict(torch.load(weights, map_location=device)['model'])
else: # darknet format
_ = load_darknet_weights(model, weights)
# Eval mode
model.to(device).eval()
# Half precision
opt.half = opt.half and device.type != 'cpu' # half precision only supported on CUDA
if opt.half:
model.half()
# Set Dataloader
vid_path, vid_writer = None, None
if opt.webcam:
save_images = False
dataloader = LoadWebcam(img_size=img_size, half=opt.half)
else:
dataloader = LoadImages(images, img_size=img_size, half=opt.half)
# Get classes and colors
# parse_data_cfg(data)['names']:得到类别名称文件路径 names=data/coco.names
classes = load_classes(parse_data_cfg(data)['names']) # 得到类别名列表: ['person', 'bicycle'...]
colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(len(classes))] # 对于每种类别随机使用一种颜色画框
# Run inference
t0 = time.time()
for i, (path, img, im0, vid_cap) in enumerate(dataloader):
t = time.time()
# if i < 500 or i % 5 == 0:
# continue
save_path = str(Path(output) / Path(path).name) # 保存的路径
# Get detections shape: (3, 416, 320)
img = torch.from_numpy(img).unsqueeze(0).to(device) # torch.Size([1, 3, 416, 320])
pred, _ = model(img) # 经过处理的网络预测，和原始的
det = non_max_suppression(pred.float(), conf_thres, nms_thres)[0] # torch.Size([5, 7])
if det is not None and len(det) > 0:
# Rescale boxes from 416 to true image size 映射到原图
det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()
# Print results to screen image 1/3 data\samples\000493.jpg: 288x416 5 persons, Done. (0.869s)
print('%gx%g ' % img.shape[2:], end='') # print image size '288x416'
for c in det[:, -1].unique(): # 对图片的所有类进行遍历循环
n = (det[:, -1] == c).sum() # 得到了当前类别的个数，也可以用来统计数目
if classes[int(c)] == 'person':
print('%g %ss' % (n, classes[int(c)]), end=', ') # 打印个数和类别'5 persons'
# Draw bounding boxes and labels of detections
# (x1y1x2y2, obj_conf, class_conf, class_pred)
count = 0
gallery_img = []
gallery_loc = []
for *xyxy, conf, cls_conf, cls in det: # 对于最后的预测框进行遍历
# *xyxy: 对于原图来说的左上角右下角坐标: [tensor(349.), tensor(26.), tensor(468.), tensor(341.)]
if save_txt: # Write to file
with open(save_path + '.txt', 'a') as file:
file.write(('%g ' * 6 + '\n') % (*xyxy, cls, conf))
# Add bbox to the image
label = '%s %.2f' % (classes[int(cls)], conf) # 'person 1.00'
if classes[int(cls)] == 'person':
#plot_one_bo x(xyxy, im0, label=label, color=colors[int(cls)])
xmin = int(xyxy[0])
ymin = int(xyxy[1])
xmax = int(xyxy[2])
ymax = int(xyxy[3])
w = xmax - xmin # 233
h = ymax - ymin # 602
# 如果检测到的行人太小了，感觉意义也不大
# 这里需要根据实际情况稍微设置下
if w*h > 100:
gallery_loc.append((xmin, ymin, xmax, ymax))
crop_img = im0[ymin:ymax, xmin:xmax] # HWC (602, 233, 3)
crop_img = Image.fromarray(cv2.cvtColor(crop_img, cv2.COLOR_BGR2RGB)) # PIL: (233, 602)
crop_img = build_transforms(reidCfg)(crop_img).unsqueeze(0) # torch.Size([1, 3, 256, 128])
gallery_img.append(crop_img)
if gallery_img:
gallery_img = torch.cat(gallery_img, dim=0) # torch.Size([7, 3, 256, 128])
gallery_img = gallery_img.to(device)
gallery_feats = reidModel(gallery_img) # torch.Size([7, 2048])
print("The gallery feature is normalized")
gallery_feats = torch.nn.functional.normalize(gallery_feats, dim=1, p=2) # 计算出查询图片的特征向量
# m: 2
# n: 7
m, n = query_feats.shape[0], gallery_feats.shape[0]
distmat = torch.pow(query_feats, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gallery_feats, 2).sum(dim=1, keepdim=True).expand(n, m).t()
# out=(beta∗M)+(alpha∗mat1@mat2)
# qf^2 + gf^2 - 2 * qf@gf.t()
# distmat - 2 * qf@gf.t()
# distmat: qf^2 + gf^2
# qf: torch.Size([2, 2048])
# gf: torch.Size([7, 2048])
distmat.addmm_(1, -2, query_feats, gallery_feats.t())
# distmat = (qf - gf)^2
# distmat = np.array([[1.79536, 2.00926, 0.52790, 1.98851, 2.15138, 1.75929, 1.99410],
# [1.78843, 1.96036, 0.53674, 1.98929, 1.99490, 1.84878, 1.98575]])
distmat = distmat.cpu().numpy() # <class 'tuple'>: (3, 12)
distmat = distmat.sum(axis=0) / len(query_feats) # 平均一下query中同一行人的多个结果
index = distmat.argmin()
if distmat[index] < dist_thres:
print('距离：%s'%distmat[index])
plot_one_box(gallery_loc[index], im0, label='(#-(T)-)', color=colors[1])#int(cls)
# cv2.imshow('person search', im0)
# cv2.waitKey(0)
print('Done. (%.3fs)' % (time.time() - t))
if opt.webcam: # Show live webcam
cv2.imshow(weights, im0)
if save_images: # Save image with detections
if dataloader.mode == 'images':
cv2.imwrite(save_path, im0)
else:
if vid_path != save_path: # new video
vid_path = save_path
if isinstance(vid_writer, cv2.VideoWriter):
vid_writer.release() # release previous video writer
fps = vid_cap.get(cv2.CAP_PROP_FPS)
width = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*fourcc), fps, (width, height))
vid_writer.write(im0)
if save_images:
print('Results saved to %s' % os.getcwd() + os.sep + output)
if platform == 'darwin': # macos
os.system('open ' + output + ' ' + save_path)
print('Done. (%.3fs)' % (time.time() - t0))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--cfg', type=str, default='cfg/yolov3.cfg', help="模型配置文件路径")
parser.add_argument('--data', type=str, default='data/coco.data', help="数据集配置文件所在路径")
parser.add_argument('--weights', type=str, default='weights/yolov3.weights', help='模型权重文件路径')
parser.add_argument('--images', type=str, default='data/samples', help='需要进行检测的图片文件夹')
parser.add_argument('-q', '--query', default=r'query', help='查询图片的读取路径.')
parser.add_argument('--img-size', type=int, default=416, help='输入分辨率大小')
parser.add_argument('--conf-thres', type=float, default=0.1, help='物体置信度阈值')
parser.add_argument('--nms-thres', type=float, default=0.15, help='NMS阈值')
parser.add_argument('--dist_thres', type=float, default=1.0, help='行人图片距离阈值，小于这个距离，就认为是该行人')
parser.add_argument('--fourcc', type=str, default='mp4v', help='fourcc output video codec (verify ffmpeg support)')
parser.add_argument('--output', type=str, default='output', help='检测后的图片或视频保存的路径')
parser.add_argument('--half', default=False, help='是否采用半精度FP16进行推理')
parser.add_argument('--webcam', default=False, help='是否使用摄像头进行检测')
opt = parser.parse_args()
print(opt)
with torch.no_grad():
detect(opt.cfg,
opt.data,
opt.weights,
images=opt.images,
img_size=opt.img_size,
conf_thres=opt.conf_thres,
nms_thres=opt.nms_thres,
dist_thres=opt.dist_thres,
fourcc=opt.fourcc,
output=opt.output)
输出，可以看到速度还挺快，yolov3能做到几乎实时：
video 1/1 (6335/6341) data\samples\badminton_mot.mp4: 256x416 6 persons, The gallery feature is normalized
Done. (0.055s)
video 1/1 (6336/6341) data\samples\badminton_mot.mp4: 256x416 5 persons, The gallery feature is normalized
Done. (0.047s)
video 1/1 (6337/6341) data\samples\badminton_mot.mp4: 256x416 4 persons, The gallery feature is normalized
Done. (0.050s)
video 1/1 (6338/6341) data\samples\badminton_mot.mp4: 256x416 5 persons, The gallery feature is normalized
Done. (0.045s)
video 1/1 (6339/6341) data\samples\badminton_mot.mp4: 256x416 5 persons, The gallery feature is normalized
Done. (0.046s)
video 1/1 (6340/6341) data\samples\badminton_mot.mp4: 256x416 5 persons, The gallery feature is normalized
Done. (0.052s)
video 1/1 (6341/6341) data\samples\badminton_mot.mp4: 256x416 5 persons, The gallery feature is normalized
Done. (0.045s)
Results saved to E:\sth22_reid\person_search_demo-master\output
Done. (371.999s)
结果动图：
2643
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
