想把计算一个人的相似度展成18个进程肯定不行。
代码只进行了18次循环处理俩个人18个关键点的距离。单进程耗时0.001秒，多进程耗时6.34秒。
import cv2
import numpy as np
from modules.keypoints import BODY_PARTS_KPT_IDS, BODY_PARTS_PAF_IDS
from modules.one_euro_filter import OneEuroFilter
import pandas as pd
class Pose:
num_kpts = 18
kpt_names = ['nose', 'neck',
'r_sho', 'r_elb', 'r_wri', 'l_sho', 'l_elb', 'l_wri',
'r_hip', 'r_knee', 'r_ank', 'l_hip', 'l_knee', 'l_ank',
'r_eye', 'l_eye',
'r_ear', 'l_ear']
sigmas = np.array([.26, .79, .79, .72, .62, .79, .72, .62, 1.07, .87, .89, 1.07, .87, .89, .25, .25, .35, .35],
dtype=np.float32) / 10.0
vars = (sigmas * 2) ** 2
last_id = -1
color = [0, 224, 255]
def __init__(self, keypoints, confidence):
super().__init__()
self.all_save_image=0
self.keypoints = keypoints
self.confidence = confidence
self.bbox = Pose.get_bbox(self.keypoints)
self.id = None
self.filters = [[OneEuroFilter(), OneEuroFilter()] for _ in range(Pose.num_kpts)]
@staticmethod
def get_bbox(keypoints):
found_keypoints = np.zeros((np.count_nonzero(keypoints[:, 0] != -1), 2), dtype=np.int32)
found_kpt_id = 0
for kpt_id in range(Pose.num_kpts):
if keypoints[kpt_id, 0] == -1:
continue
found_keypoints[found_kpt_id] = keypoints[kpt_id]
found_kpt_id += 1
bbox = cv2.boundingRect(found_keypoints)
return bbox
def update_id(self, id=None):
self.id = id
if self.id is None:
self.id = Pose.last_id + 1
Pose.last_id += 1
def update_all(self, id=None):
self.all_save_image = id
if self.all_save_image is None:
self.all_save_image = 0
def draw(self, img):
assert self.keypoints.shape == (Pose.num_kpts, 2)
for part_id in range(len(BODY_PARTS_PAF_IDS) - 2):
kpt_a_id = BODY_PARTS_KPT_IDS[part_id][0]
global_kpt_a_id = self.keypoints[kpt_a_id, 0]
if global_kpt_a_id != -1:
x_a, y_a = self.keypoints[kpt_a_id]
cv2.circle(img, (int(x_a), int(y_a)), 3, Pose.color, -1)
kpt_b_id = BODY_PARTS_KPT_IDS[part_id][1]
global_kpt_b_id = self.keypoints[kpt_b_id, 0]
if global_kpt_b_id != -1:
x_b, y_b = self.keypoints[kpt_b_id]
cv2.circle(img, (int(x_b), int(y_b)), 3, Pose.color, -1)
if global_kpt_a_id != -1 and global_kpt_b_id != -1:
cv2.line(img, (int(x_a), int(y_a)), (int(x_b), int(y_b)), Pose.color, 2)
def get_similarity(a, b, threshold=0.5):
num_similar_kpt = 0
for kpt_id in range(Pose.num_kpts):
if a.keypoints[kpt_id, 0] != -1 and b.keypoints[kpt_id, 0] != -1:
distance = np.sum((a.keypoints[kpt_id] - b.keypoints[kpt_id]) ** 2)
area = max(a.bbox[2] * a.bbox[3], b.bbox[2] * b.bbox[3])
similarity = np.exp(-distance / (2 * (area + np.spacing(1)) * Pose.vars[kpt_id]))
if similarity > threshold:
num_similar_kpt += 1
return num_similar_kpt
import time
import numpy as np
from concurrent.futures import ProcessPoolExecutor
def s(kpt_id,threshold=0.5):
if a.keypoints[kpt_id, 0] != -1 and b.keypoints[kpt_id, 0] != -1:
distance = np.sum((a.keypoints[kpt_id] - b.keypoints[kpt_id]) ** 2)
area = max(a.bbox[2] * a.bbox[3], b.bbox[2] * b.bbox[3])
similarity = np.exp(-distance / (2 * (area + np.spacing(1)) * Pose.vars[kpt_id]))
if similarity > threshold:
return 1
if __name__=='__main__':
keypoint=np.random.random((18,2))
pose=Pose(keypoint,1)
a=pose
b=pose
start=time.time()
get_similarity(pose,pose)
end=time.time()
print(end-start)
ind=18
num_similar_kpt=0
start=time.time()
with ProcessPoolExecutor() as pool:
results=pool.map(s,[i for i in range(ind)])
end=time.time()
sum1=end-start
print(sum1)
