当前位置：首页 > news >正文

计算机视觉的应用17-利用CrowdCountNet模型解决人群数量计算问题(pytorch搭建模型)

news 来源：原创 2024/5/14 17:53:11

大家好，我是微学AI，今天给大家介绍一下计算机视觉的应用17-利用CrowdCountNet模型解决人群数量计算问题(pytorch搭建模型)。本篇文章，我将向大家展示如何使用CrowdCountNet这个神奇的工具，以及它是如何利用深度学习技术来解决复杂的人群计数问题。让我们一起进入这个充满活力和创新的世界，开启图像和视频中人群数量计算的新篇章！

1. 项目介绍

本文我将利用深度神经网络来解决一个现实中普遍存在的问题：如何准确计算图像和视频中的人群数量。当您走进拥挤的城市街头或繁忙的公共场所时，人群数量经常让人难以置信。然而，现在有了深度学习模型的帮助，我们可以轻松地通过计算机视觉来解决这个挑战。

CrowdCountNet是我们的主角，它是一种被广泛应用于图像识别和处理领域的深度学习模型。它背后的原理十分精巧，利用了神经网络的强大能力来理解和分析图像中的人群分布。这个模型通过学习大量的图像数据，自动捕捉到了各种人群密集度的模式和特征。

想象一下，当你看着一张摄像头拍摄的城市街景时，CrowdCountNet正在忙碌地工作着。它会逐像素地扫描整个图像，并识别每个像素点上是否存在人群。从细微的行人到人群聚集的区域，CrowdCountNet都能准确地捕捉到每个人的存在。

使用这个强大的深度学习模型，我们可以实现许多令人惊叹的功能。无论是为城市规划提供人流热图、帮助安保人员监控拥挤场所，还是为交通管理提供实时的交通流量信息，CrowdCountNet都能在不同领域发挥巨大作用。

2. 应用场景

2.1 人流监测和管理

在公共场所，例如商场、机场、火车站等，监测和管理人流量是至关重要的。我们的模型可以用于实时监测人流量，帮助管理者做出更有效的决策，比如调整人流方向，预防拥挤等。

在一个繁忙的购物中心，通过我们的人流监测系统，可以实时显示各个商店的人流量。商场管理员可以在控制中心的大屏幕上看到不同区域的人流状况，比如一楼的时尚区人流量饱和，而二楼的电子产品区人流相对稀少。管理员立即作出反应，调整楼梯和电梯的方向，引导顾客流向较空闲的区域，以缓解拥挤。

2.2 安全防控

在大型活动或集会中，通过实时监测人群数量，可以预防和控制安全事故的发生，及时制定疏散计划，提高人员安全。

想象一个音乐节现场，数以万计的观众聚集在一个开放的场地上。通过我们的人流监测系统，主办方能够实时获得观众的数量和密度数据。突然，系统发出警报，显示某个区域的人流超过了安全限制。主办方立刻启动紧急预案，引导人群有序撤离，避免发生踩踏事故。

2.3 市场调研和决策支持

商家可以通过监测店铺或某个区域的人流量，来评估其营销策略的效果，或者进行更准确的市场调研。

一家新开业的百货公司想要评估其广告宣传效果和吸引力。通过人流监测系统，他们可以统计每天进入商场的人数，并与营销活动的时间和内容进行对比。他们发现，当进行打折促销时，入场人数骤增，而在没有促销的日子里，人流量相对稀少。这为他们提供了有价值的市场调研数据，帮助他们更准确地评估促销策略的成效。

2.4 城市规划和交通管理

在城市规划和交通管理中，通过人群数量的监测，可以更好地理解和预测城市中的人流动态，从而更科学地进行城市规划和交通管理。

想象一座拥挤的大都市，上班高峰期大批人涌入地铁站。通过我们的人流监测系统，地铁管理部门可以实时了解不同地铁站的客流情况，并根据需求增加或减少列车班次。当一个地铁站即将达到容纳上限时，系统会自动发出警报，引导乘客选择其他线路或利用公共交通换乘，以减少人流压力。

2.5 相册与毕业生人数统计

通常毕业合影中有大批的人一起合作，我们要统计人数的话，基本都是人工去一个一个数出来，这样费时费力。现在利用模型直接统计合照人数，统计合照人数，判断是否学生来齐了。

3. 实战项目

在这个项目中，我们将首先讨论如何准备数据，然后构建和训练我们的模型，并最后使用我们的模型来检测人群数量。

3.1 数据准备

首先，我们需要准备一个包含大量标注人群数量的图像数据集。这可以是公开的人群数量数据集，也可以是自己收集并标注的数据集。

3.2 模型构建

然后，我们需要构建一个CrowdCountNet模型来学习如何从图像中预测人群数量。

import torch.nn as nn
import torch
import torch.nn.functional as F
from torch.autograd import Variable
from torch.nn.utils.weight_norm import weight_norm
import math
from collections import OrderedDictclass CrowdCountNet(nn.Module):def __init__(self,leaky_relu=False,attn_weight=1,fix_domain=1,domain_center_model='',**kwargs):super(CrowdCountNet, self).__init__()self.criterion_attn = torch.nn.MSELoss(reduction='sum')self.domain_center_model = domain_center_modelself.attn_weight = attn_weightself.fix_domain = fix_domainself.cosine = 1self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=2, padding=1, bias=False)self.bn1 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM)self.conv2 = nn.Conv2d(64, 64, kernel_size=3, stride=2, padding=1, bias=False)self.bn2 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM)self.relu = nn.ReLU(inplace=True)num_channels = 64block = blocks_dict['BOTTLENECK']num_blocks = 4self.layer1 = self._make_layer(block, 64, num_channels, num_blocks)stage1_out_channel = block.expansion * num_channels# -- stage 2self.stage2_cfg = {}self.stage2_cfg['NUM_MODULES'] = 1self.stage2_cfg['NUM_BRANCHES'] = 2self.stage2_cfg['BLOCK'] = 'BASIC'self.stage2_cfg['NUM_BLOCKS'] = [4, 4]self.stage2_cfg['NUM_CHANNELS'] = [40, 80]self.stage2_cfg['FUSE_METHOD'] = 'SUM'num_channels = self.stage2_cfg['NUM_CHANNELS']block = blocks_dict[self.stage2_cfg['BLOCK']]num_channels = [num_channels[i] * block.expansionfor i in range(len(num_channels))]self.transition1 = self._make_transition_layer([stage1_out_channel],num_channels)self.stage2, pre_stage_channels = self._make_stage(self.stage2_cfg, num_channels)# -- stage 3self.stage3_cfg = {}self.stage3_cfg['NUM_MODULES'] = 4self.stage3_cfg['NUM_BRANCHES'] = 3self.stage3_cfg['BLOCK'] = 'BASIC'self.stage3_cfg['NUM_BLOCKS'] = [4, 4, 4]self.stage3_cfg['NUM_CHANNELS'] = [40, 80, 160]self.stage3_cfg['FUSE_METHOD'] = 'SUM'num_channels = self.stage3_cfg['NUM_CHANNELS']block = blocks_dict[self.stage3_cfg['BLOCK']]num_channels = [num_channels[i] * block.expansionfor i in range(len(num_channels))]self.transition2 = self._make_transition_layer(pre_stage_channels,num_channels)self.stage3, pre_stage_channels = self._make_stage(self.stage3_cfg, num_channels)last_inp_channels = np.int(np.sum(pre_stage_channels)) + 256self.redc_layer = nn.Sequential(nn.Conv2d(in_channels=last_inp_channels,out_channels=128,kernel_size=3,stride=1,padding=1),nn.BatchNorm2d(128, momentum=BN_MOMENTUM),nn.ReLU(True),)self.aspp = nn.ModuleList(aspp(in_channel=128))# additional layers specfic for Phase 3self.pred_conv = nn.Conv2d(128, 512, 3, padding=1)self.pred_bn = nn.BatchNorm2d(512)self.GAP = nn.AdaptiveAvgPool2d(1)# Specially for hidden domain# Set the domain for learnable parametersdomain_center_src = np.load(self.domain_center_model)G_SHA = torch.from_numpy(domain_center_src['G_SHA']).view(1, -1, 1, 1)G_SHB = torch.from_numpy(domain_center_src['G_SHB']).view(1, -1, 1, 1)G_QNRF = torch.from_numpy(domain_center_src['G_QNRF']).view(1, -1, 1, 1)self.n_domain = 3self.G_all = torch.cat([G_SHA.clone(), G_SHB.clone(),G_QNRF.clone()], dim=0)self.G_all = nn.Parameter(self.G_all)self.last_layer = nn.Sequential(nn.Conv2d(in_channels=128,out_channels=64,kernel_size=3,stride=1,padding=1),nn.BatchNorm2d(64, momentum=BN_MOMENTUM),nn.ReLU(True),nn.Conv2d(in_channels=64,out_channels=32,kernel_size=3,stride=1,padding=1),nn.BatchNorm2d(32, momentum=BN_MOMENTUM),nn.ReLU(True),nn.Conv2d(in_channels=32,out_channels=1,kernel_size=1,stride=1,padding=0),)def _make_transition_layer(self, num_channels_pre_layer,num_channels_cur_layer):num_branches_cur = len(num_channels_cur_layer)num_branches_pre = len(num_channels_pre_layer)transition_layers = []for i in range(num_branches_cur):if i < num_branches_pre:if num_channels_cur_layer[i] != num_channels_pre_layer[i]:transition_layers.append(nn.Sequential(nn.Conv2d(num_channels_pre_layer[i],num_channels_cur_layer[i],3,1,1,bias=False),nn.BatchNorm2d(num_channels_cur_layer[i],momentum=BN_MOMENTUM), nn.ReLU(inplace=True)))else:transition_layers.append(None)else:conv3x3s = []for j in range(i + 1 - num_branches_pre):inchannels = num_channels_pre_layer[-1]outchannels = num_channels_cur_layer[i] \if j == i - num_branches_pre else inchannelsconv3x3s.append(nn.Sequential(nn.Conv2d(inchannels, outchannels, 3, 2, 1, bias=False),nn.BatchNorm2d(outchannels, momentum=BN_MOMENTUM),nn.ReLU(inplace=True)))transition_layers.append(nn.Sequential(*conv3x3s))return nn.ModuleList(transition_layers)def _make_layer(self, block, inplanes, planes, blocks, stride=1):downsample = Noneif stride != 1 or inplanes != planes * block.expansion:downsample = nn.Sequential(nn.Conv2d(inplanes,planes * block.expansion,kernel_size=1,stride=stride,bias=False),nn.BatchNorm2d(planes * block.expansion, momentum=BN_MOMENTUM),)layers = []layers.append(block(inplanes, planes, stride, downsample))inplanes = planes * block.expansionfor i in range(1, blocks):layers.append(block(inplanes, planes))return nn.Sequential(*layers)def _make_stage(self,layer_config,num_inchannels,multi_scale_output=True):num_modules = layer_config['NUM_MODULES']num_branches = layer_config['NUM_BRANCHES']num_blocks = layer_config['NUM_BLOCKS']num_channels = layer_config['NUM_CHANNELS']block = blocks_dict[layer_config['BLOCK']]fuse_method = layer_config['FUSE_METHOD']modules = []for i in range(num_modules):# multi_scale_output is only used last moduleif not multi_scale_output and i == num_modules - 1:reset_multi_scale_output = Falseelse:reset_multi_scale_output = Truemodules.append(HighResolutionModule(num_branches, block, num_blocks,num_inchannels, num_channels, fuse_method,reset_multi_scale_output))num_inchannels = modules[-1].get_num_inchannels()return nn.Sequential(*modules), num_inchannelsdef forward(self, x):x = self.conv1(x)x = self.bn1(x)x = self.relu(x)x = self.conv2(x)x = self.bn2(x)x = self.relu(x)x = self.layer1(x)x_head_1 = xx_list = []for i in range(self.stage2_cfg['NUM_BRANCHES']):if self.transition1[i] is not None:x_list.append(self.transition1[i](x))else:x_list.append(x)y_list = self.stage2(x_list)x_list = []for i in range(self.stage3_cfg['NUM_BRANCHES']):if self.transition2[i] is not None:x_list.append(self.transition2[i](y_list[-1]))else:x_list.append(y_list[i])x = self.stage3(x_list)# Replace the classification heaeder with custom setting# Upsamplingx0_h, x0_w = x[0].size(2), x[0].size(3)x1 = F.interpolate(x[1], size=(x0_h, x0_w), mode='bilinear', align_corners=False)x2 = F.interpolate(x[2], size=(x0_h, x0_w), mode='bilinear', align_corners=False)x = torch.cat([x[0], x1, x2, x_head_1], 1)# first, reduce the channel downx = self.redc_layer(x)pred_attn = self.GAP(F.relu_(self.pred_bn(self.pred_conv(x))))pred_attn = F.softmax(pred_attn, dim=1)pred_attn_list = torch.chunk(pred_attn, 4, dim=1)aspp_out = []for k, v in enumerate(self.aspp):if k % 2 == 0:aspp_out.append(self.aspp[k + 1](v(x)))else:continue# Using Aspp add, and relu insidefor i in range(4):x = x + F.relu_(aspp_out[i] * 0.25) * pred_attn_list[i]bz = x.size(0)# -- Besides, we also need to let the prediction attention be close to visable domain# -- Calculate the domain distance and get the weights# - First, detach domainsG_all_d = self.G_all.detach()  # use detached G_all for calulcatingpred_attn_d = pred_attn.detach().view(bz, 512, 1, 1)if self.cosine == 1:G_A, G_B, G_Q = torch.chunk(G_all_d, self.n_domain, dim=0)cos_dis_A = F.cosine_similarity(pred_attn_d, G_A, dim=1).view(-1)cos_dis_B = F.cosine_similarity(pred_attn_d, G_B, dim=1).view(-1)cos_dis_Q = F.cosine_similarity(pred_attn_d, G_Q, dim=1).view(-1)cos_dis_all = torch.stack([cos_dis_A, cos_dis_B,cos_dis_Q]).view(bz, -1)  # bz*3cos_dis_all = F.softmax(cos_dis_all, dim=1)target_attn = cos_dis_all.view(bz, self.n_domain, 1, 1, 1).expand(bz, self.n_domain, 512, 1, 1) * self.G_all.view(1, self.n_domain, 512, 1, 1).expand(bz, self.n_domain, 512, 1, 1)target_attn = torch.sum(target_attn, dim=1, keepdim=False)  # bz * 512 * 1 * 1if self.fix_domain:target_attn = target_attn.detach()else:raise ValueError('Have not implemented not cosine distance yet')x = self.last_layer(x)x = F.relu_(x)x = F.interpolate(x, size=(x0_h * 2, x0_w * 2), mode='bilinear', align_corners=False)return x, pred_attn, target_attndef init_weights(self,pretrained='',):logger.info('=> init weights from normal distribution')for m in self.modules():if isinstance(m, nn.Conv2d):nn.init.normal_(m.weight, std=0.01)if m.bias is not None:nn.init.constant_(m.bias, 0)elif isinstance(m, nn.BatchNorm2d):nn.init.constant_(m.weight, 1)nn.init.constant_(m.bias, 0)if os.path.isfile(pretrained):pretrained_dict = torch.load(pretrained)logger.info(f'=> loading pretrained model {pretrained}')model_dict = self.state_dict()pretrained_dict = {k: vfor k, v in pretrained_dict.items() if k in model_dict.keys()}for k, _ in pretrained_dict.items():logger.info(f'=> loading {k} pretrained model {pretrained}')model_dict.update(pretrained_dict)self.load_state_dict(model_dict)else:assert 1 == 2

3.3 图片检测人群数量

接下来，我们加载已经训练好的模型进行预测，这里省略了中间复杂的过程，大家可以一键调用预测。

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
from modelscope.outputs import OutputKeys
from modelscope.utils.cv.image_utils import numpy_to_cv2img
import cv2crowd_model = pipeline(Tasks.crowd_counting,model='damo/cv_hrnet_crowd-counting_dcanet')imgs = '111.png'
results = crowd_model(imgs)print('人数为：', int(results[OutputKeys.SCORES]))vis_img = results[OutputKeys.OUTPUT_IMG]
vis_img = numpy_to_cv2img(vis_img)
cv2.imwrite('result1.jpg', vis_img)

在这里插入图片描述
预测生成图：

毕业合照人群计数：

预测生成图：

3.4 视频检测人群数量

对于视频，我们可以将其分解为一系列的图像帧，然后使用我们的模型来检测每一帧中的人群数量。

import cv2
from modelscope.outputs import OutputKeysdef predict_video(video_path):cap = cv2.VideoCapture(video_path)while (cap.isOpened()):ret, frame = cap.read()if ret:gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)preds = crowd_model(gray)print(preds[OutputKeys.SCORES])else:breakcap.release()video_path = 'test.mp4'
predict_video(video_path)