利用pycharm阅读代码,进行Debug
objdetector.py 注释
import torch
import numpy as np
from models.experimental import attempt_load
from utils.general import non_max_suppression, scale_coords
from utils.datasets import letterbox
from utils.torch_utils import select_device
import objtracker
# 要检测的类别,这里只检测人,车(小车、巴士、卡车)
OBJ_LIST = ['person', 'car', 'bus', 'truck']
# yolov5模型
DETECTOR_PATH = 'weights/yolov5m.pt'
class baseDet(object):
def __init__(self):
self.img_size = 640
self.threshold = 0.3
self.stride = 1
def build_config(self):
# 帧数统计变量初始化
self.frameCounter = 0
def feedCap(self, im, func_status):
# 初始化dict,用于返回结果, 在这里初始化了键list_of_ids, 但后面没用到
retDict = {
'frame': None,
'list_of_ids': None,
'obj_bboxes': []
}
# 帧数统计
self.frameCounter += 1
# feed to 检测器+deepsort, 注意第一个参数self《=》det. 返回带绘制检测+deepsort结果的im,和检测+deepsort结果信息:[(x1, y1, x2, y2, '', track_id),...,(...)]
im, obj_bboxes = objtracker.update(self, im)
retDict['frame'] = im
retDict['obj_bboxes'] = obj_bboxes
return retDict
def init_model(self):
raise EOFError("Undefined model type.")
def preprocess(self):
raise EOFError("Undefined model type.")
def detect(self):
raise EOFError("Undefined model type.")
# 定义类Detector,继承自baseDet
class Detector(baseDet):
# 构造函数
def __init__(self):
# 调用父类构造函数初始化继承自父类的属性,具体是父类构造函数中定义的属性
super(Detector, self).__init__()
# 调用init_model方法
self.init_model()
# 调用父类方法
self.build_config()
def init_model(self):
# 权重文件 yolov5s.pt
self.weights = DETECTOR_PATH
# 选择使用gpu or cpu,我这里会使用cpu
self.device = '0' if torch.cuda.is_available() else 'cpu'
# yolov5中提供的select_device方法直接复用
self.device = select_device(self.device)
# 加载权重文件
model = attempt_load(self.weights, map_location=self.device)
# 将model transfer to device, 推理阶段使用model.eval() 训练阶段使用model.train()
model.to(self.device).eval()
# GPU支持半精度
#model.half()
# 使用cpu时不支持.half,改为.float(), 这里比较奇怪,我明明有gpu但选择的cpu,后面再细究
model.float()
self.m = model
# 模型能够检测的所有类别标签。yolov5中提供的方法直接复用
self.names = model.module.names if hasattr(
model, 'module') else model.names
def preprocess(self, img):
# 对原图进行拷贝
img0 = img.copy()
# yolov5图像预处理之lettexbox, [1080,1920,3]->[384,640,3]
img = letterbox(img, new_shape=self.img_size)[0]
# [384,640,3] -> [3,384,640]
img = img[:, :, ::-1].transpose(2, 0, 1)
# 在内存上使用连续的内存存储图像
img = np.ascontiguousarray(img)
# 由numpy array创建torch Tensor; transfer to device
img = torch.from_numpy(img).to(self.device)
#img = img.half() # 半精度
img = img.float()
img /= 255.0 # 图像归一化
# [3,384,640] -> [1,3,384,640]
if img.ndimension() == 3:
img = img.unsqueeze(0)
# 返回原图像img0和预处理之后的图像img
return img0, img
def detect(self, im):
# 图像预处理,返回原图im0,预处理之后的图像img
im0, img = self.preprocess(im)
# feed to yolov5 进行检测,调用yolo.py Class Model中的forward方法。返回检测结果
pred = self.m(img, augment=False)[0]
pred = pred.float()
# NMS
pred = non_max_suppression(pred, self.threshold, 0.4)
# 初始化返回结果
pred_boxes = []
# yolov5检测结果解析,yolov5源码方法复用
for det in pred:
if det is not None and len(det):
# 调整检测框坐标。检测框是基于预处理后640x640的图像的,调整为基于原图的检测框
det[:, :4] = scale_coords(
img.shape[2:], det[:, :4], im0.shape).round()
for *x, conf, cls_id in det:
lbl = self.names[int(cls_id)]
# 筛选出要检测的类别,过滤掉其它类别
if not lbl in OBJ_LIST:
continue
x1, y1 = int(x[0]), int(x[1])
x2, y2 = int(x[2]), int(x[3])
# 检测框坐标(左上右下),类别标签,置信度
pred_boxes.append(
(x1, y1, x2, y2, lbl, conf))
# 返回原图,检测结果 pred_boxes:[(...),(,,,), ... ,(...)]
return im, pred_boxes