from Tracking.deep_sort.detection import Detection
from Tracking import generate_dets as gdet
from Tracking.deep_sort.tracker import Tracker
from keras.models import load_model
from .action_enum import Actions
# 기본 매개 변수 정의
file_path = Path.cwd()
clip_length = 15
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
# 초기화 deep_sort
model_filename = str(file_path / 'Tracking/graph_model/mars-small128.pb')
encoder = gdet.create_box_encoder(model_filename, batch_size=1)
metric = NearestNeighborDistanceMetric("cosine", max_cosine_distance,
nn_budget)
tracker = Tracker(metric)
# track_box 색상
trk_clr = (0, 255, 0)
def load_action_premodel(model):
return load_model(model)
def framewise_recognize(pose, pretrained_model):
frame, joints, bboxes, xcenter = pose[0], pose[1], pose[2], pose[3]
joints_norm_per_frame = np.array(pose[-1])
# Use Deep-sort(Simple Online and Realtime Tracking)
# To track multi-person for multi-person actions recognition
# 定义基本参数
file_path = Path.cwd()
clip_length = 15
max_cosine_distance = 0.3
nn_budget = None
nms_max_overlap = 1.0
fall_num = 0
# 初始化deep_sort
model_filename = str(file_path / 'Tracking/graph_model/mars-small128.pb')
#对检测到的object path 编码
encoder = gdet.create_box_encoder(model_filename,
batch_size=1) #encoder的索引 计算得到特征 对特征进行编码
#度量kalman预测的目标和下一帧的检测目标进行距离计算 使用余弦距离能够缓解遮挡 ID switch比较频繁的问题
metric = NearestNeighborDistanceMetric("cosine", max_cosine_distance,
nn_budget) #
tracker = Tracker(metric) #根据度量结果追踪
# track_box颜色
trk_clr = (0, 255, 0)
# class ActionRecognizer(object):
# @staticmethod
# def load_action_premodel(model):
# return load_model(model)
#
# @staticmethod
# def framewise_recognize(pose, pretrained_model):
