CCTwins: A Weakly-Supervised Transformer-based Crowd Counting Method with Adaptive Scene Consistency Attention

Recently, crowd counting has attracted significant attention, particularly in the context of the COVID-19 pandemic, due to its ability to automatically provide accurate crowd numbers in images. To address the challenges of location-level labeling, several transformer-based crowd counting methods have been proposed with only count-level supervision. However, these methods directly use the transformer as an encoder without considering the uneven crowd distribution. To address this issue, we propose CCTwins, a novel transformer-based crowd counting method with only count-level supervision. Specifically, we introduce an adaptive scene consistency attention mechanism to enhance the transformer-based model Twins-SVT-L for feature extraction in crowded scenes. Additionally, we design a multi-level weakly-supervised loss function that generates estimated crowd numbers in a coarse-to-fine manner, making it more appropriate for weakly-supervised settings. Moreover, intermediate features supervised by count-level labels are utilized to fuse multi-scale features. Experimental results on four public datasets demonstrate that our proposed method outperforms the state-of-the-art weakly-supervised methods, achieving up to a 16.6% improvement in MAE and up to a 13.8% improvement in RMSE across all evaluation settings. Moreover, the proposed CCTwins obtains competitive counting performance, even when compared to the state-of-the-art fully-supervised methods. IEEE

Cascaded parallel crowd counting network with multi-resolution collaborative representation.

Accurately estimating the size and density distribution of a crowd from images is of great importance to public safety and crowd management during the COVID-19 pandemic, but it is very challenging as it is affected by many complex factors, including perspective distortion and background noise information. In this paper, we propose a novel multi-resolution collaborative representation framework called the cascaded parallel network (CP-Net), consisting of three parallel scale-specific branches connected in a cascading mode. In the framework, the three cascaded multi-resolution branches efficiently capture multi-scale features through their specific receptive fields. Additionally, multi-level feature fusion and information filtering are performed continuously on each branch to resist noise interference and perspective distortion. Moreover, we design an information exchange module across independent branches to refine the features extracted by each specific branch and deal with perspective distortion by using complementary information of multiple resolutions. To further improve the robustness of the network to scale variance and generate high-quality density maps, we construct a multi-receptive field fusion module to aggregate multi-scale features more comprehensively. The performance of our proposed CP-Net is verified on the challenging counting datasets (UCF_CC_50, UCF-QNRF, Shanghai Tech A&B, and WorldExpo'10), and the experimental results demonstrate the superiority of the proposed method.

Crowd Counting in Public Places Using MultiScale Convolutional Neural Network

Crowd of people in public places is a serious problem that needs attention because uncontrolled crowd conditions will cause problems, especially with the Covidl9 pandemic which requires people not to congregate. This research uses the Multi Scale Convolutional Neural Network method to overcome the main problems in crowd images, namely object scale variations, difficulty distinguishing between people objects and the background, as well as overlapping between people objects. The Multi Scale CNN implementation in this research uses the feature extractor layer from VGG16 as the low level feature extractor layer (frontend layer) and the Inception-Restnet-A module from Inception-Resnet-v2 as the high level feature extractor (backend layer). The datasets used to train the model are the ShanghaiTech and UCF_QNRF datasets which already contain the location information of the people in the image. Prior to the training process, ground-truth was made by conducting a convolution process using a Gaussian filter at the point where people are. Then, the Multi Scale CNN model will be trained with these 2 datasets. In the trained model, the input image will be convoluted to produce a density map. The results of the crowd calculation are obtained by adding up all the density map values. The use of Multi Scale CNN is proven to provide a good accuracy value with the MAE loss value being 78.0 for the ShanghaiTech Part A dataset and 10.75 for the ShanghaiTech Part B dataset. © 2022 IEEE.

WiFi-Crowd Spy: A novel crowd-counting system

Currently, there is a requirement in many countries to keep public and work spaces safe due to COVID-19. In fact, indoor spaces must be monitored to control the allowed capacity, which can vary depending on the alert level of a city at a given time. This has motivated some researchers to investigate several technologies to implement methods and strategies to enable the reopening of these spaces in a safe manner. In this paper, we propose a crowd counting detection system that this paper, we propose a crowd counting detection system that addresses the problem of controlling the indoor capacity of offices inside buildings. The proposed solution uses an existing communication technology such as WiFi in order to determine the crowd counting for the indoor environment. In particular, the existing infrastructure consists of two Wireless LAN Controllers (WLC) and several APs deployed in a building, which allows us to estimate the number of people based on the access to Wireless Access Points (APs). Thus, the proposed system takes into account when a mobile device connects/disconnects to the AP to increase or decrease the number of people in a particular office and sends the respective alert to the system administrator when this capacity is about to be exceeded or already surpassed. © 2022 IEEE.

Crowd Counting Using Region Convolutional Neural Networks

Monitoring the number of people is essential to estimate the level of crowds in a public area, especially during this Covid19 pandemic. CCTV recording needs to process for counting the number of people in a crowd at a specific time. However, counting people on CCTV is not easy. It can be approached by detecting a specific object from a compilation of frames with a certain size that makes up the image. This study proposed the Faster Region-Convolutional Neural Networks (Faster R-CNN) method with ResNet50 to count the number of people in a crowd from the low-resolution image from CCTV. The research gave that crowd counting with the Faster RCNN needs consideration to choose appropriate architecture. ResNet50 architecture provided an accuracy of 97.20% in detecting the number of people in the crowd image. It was compared to other detectors based on previous studies with the same dataset and gave the highest accuracy. Region Proposal Networks makes Faster RCNN robust. Although the various number of people in the crowd image, quality of the dataset, and anchor aspect ratio values provide good results improve accuracy. Besides, the appropriate learning parameters make the method performance more optimal. This configuration can be applied to real-time testing so that it gave the best results of 86% using Faster RCNN and ResNet50.

Benchmarking Crowd-Counting Techniques across Image Characteristics

Crowd-counting is a longstanding computer vision used in estimating the crowd sizes for security purposes at public protests in streets, public gatherings, for collecting crowd statistics at airports, malls, concerts, conferences, and other similar venues, and for monitoring people and crowds during public health crises (such as the one caused by COVID-19). Recently, the performance of automated methods for crowd-counting from single images has improved particularly due to the introduction of deep learning techniques and large labelled training datasets. However, the robustness of these methods to varying imaging conditions, such as weather, image perspective, and large variations in the crowd size has not been studied in-depth in the open literature. To address this gap, a systematic study on the robustness of four recently developed crowd-counting methods is performed in this paper to evaluate their performance with respect to variable (real-life) imaging scenarios that include different event types, weather conditions, image sources and crowd sizes. It is shown that the performance of the tested techniques is degraded in unclear weather conditions (i.e., fog, rain, snow) and also on images taken from large distances by drones. On the opposite, clear weather conditions, crowd-counting methods can provide accurate and usable results. © 2021 Electrotechnical Society of Slovenia. All rights reserved.

Capacity Estimation from Environmental Audio Signals Using Deep Learning

Estimating the capacity of a room or venue is essential to avoid overcrowding that could compromise people’s safety. Having enough free space to guarantee a minimal safety distance between people is also essential for health reasons, as in the current COVID-19 pandemic. Already existing systems for automatic crowd counting are mostly based on image or video data, and some of them, using deep learning architectures. In this paper, we study the viability of already existing Deep Learning Crowd Counting systems and propose new alternatives based on new network architectures containing convolutional layers, exclusively based on the use of environmental audio signals. The proposed architecture is able to infer the actual capacity with a higher accuracy in comparison to previous proposals. Consequently, conclusions from the accuracy obtained with out approach are drawn and the possible scope of deep learning based crowd counting systems is discussed. © 2022, Springer Nature Switzerland AG.

Multi Layered Deep Neural Network for Feature Extraction in Cross Domain Crowd Counting

Automated crowd density monitoring is an emerging area of research. It is a vital technology that assists during recent disease outbreaks in preserving social distancing, crowd management and other widespread applications in public security and traffic control. Modern methods to count people in crowded scenes mainly rely on Convolutional Neural Network (CNN) based models. But the model's ability to adapt for different domains which is referred to as cross domain crowd counting is a challenging task. To remedy this difficulty, many researchers used Spatial Fully Convolutional Network (SFCN) based crowd counting models with synthetic crowd scene data. They covered many image domains with few-shot learning to reduce the domain adaptation gap between source and target image domains. In this paper, we propose a new multi-layered model architecture instead of SFCN single-layered model architecture. The proposed model extracts more meaningful features in image scenes along with large scale variations to increase the accuracy in cross domain crowd counting. Furthermore, with extensive experiments using four real-world datasets and analysis, we show that the proposed multi-layered architecture performs well with synthetic image data and few-shot learning in reducing domain shifts. © 2022 IEEE.

Social Distancing Crowd Segmentation, Estimation and Visualisation

This research shows a modern crowd counting solution which alters typical prediction solutions into a segmentation of individuals based on a distance threshold, allowing for better visualisation and results. The study proposes using YOLOv4-normal and YOLOv4-tiny models, which have shown great results throughout calibration with an MAE of 14 and 36 respectively. However it did present some issues of accuracy degradation when trained on head annotations at any level of crowd density. As for visualisation, perspective transformation was used which directly helped in providing the distance calculation that was absent from standard transformation. If any variants of YOLOv4 are to be used, the main argument is the choice between speed over accuracy while relying on native implementations. In the case of distance regulation, any transformation that maps itself onto the region of interest, such as perspective transformation should be used to precisely determine distances from a camera to the region of interest itself. © 2021 IEEE.

Integrated Video Based Crowdedness Forecasting Framework with a Review of Crowd Counting Models

Crowd counting and forecasting is an important problem amidst Covid 19 circumstances. A unified system to automate crowd monitoring, collect data about crowdedness and predict future crowds is presented in this paper. An evaluation of existing state-of-the-art crowd counting algorithms on a novel dataset is conducted in the first part of the paper, which demonstrates the shortcomings of these algorithms. Several novel algorithms, including a densely connected neural network, convolutional neural network, and a long short term memory based recurrent neural network, for predicting crowd counts in the near and distant future are presented afterwards in the second half of the paper. © 2021 IEEE.

Jointly Attention Network for Crowd Counting

Nowadays, crowd counting has shown great practical value in public safety and related fields. Most leading algorithms exploit CNN to generate density maps and have improved the estimation accuracy. However, the counting models still suffer from the challenge of huge scale variations. In order to mitigate this issue, we propose a novel approach named Jointly Attention Network (JANet) for Crowd Counting. It is composed of two major schemes: the Multi-order Scale Attention (MSA) module and the Multi-pooling Relational Channel Attention (MRCA) module. The MSA module explores meaningful high-order statistics and helps the backbone network obtain more discriminative features with rich scale information in an explicit manner. The MRCA module compactly represents the global scope relations and accounts the interdependence among all channel-wise nodes, which is complementary to MSA module. Meanwhile, the Distributed Combinatorial Loss (DCL) is designed to achieve the distributed supervision on intermediate layers at each level. Finally, we conduct extensive studies on multiple crowd counting datasets, the ShanghaiTech, the UCF-QNRF, the JHU-CROWD++, the NWPU-Crowd. The experimental results indicate that our proposed method has achieved superior performance.

Crowd counting: a behavioural economics perspective.

Developments in technology have facilitated the emergence of new crowd counting organisations. Some of the organisations have established platforms to disseminate their data, making it available to researchers for the first time. These databases promise to increase the quality and quantity of research in various fields. In the late 2010s, specialist crowd counting organisations emerged with the sole purpose of counting crowds at protests and disseminating the results, sometimes in a purely partisan manner. Because of the contemporary relevance of protest behaviour, we frame our discussion within this context. For social scientists considering the utilisation of these new databases, it is essential that crowd numbers be linked to underlying human behaviour in a way that promises a chain of connections to investigate and explore. We use behavioural economics to show why relative crowd size may be important for human decision-makers. And we show how the significance of relative crowd size relates to other aspects of the human decision-making process, including risk preferences and probability assessments. Far from being a theory of protest behaviour, we present a behavioural economics-based primer for empirical researchers and social scientists engaging with newly available crowd counting data. The conclusions may apply in other contexts and might be extended to encompass specific types of behaviour, including aggression and violence. Indeed, the conclusions may guide the analysis of the emergence of the crowd counting organisations themselves.

Deep Learning-Based Crowd Scene Analysis Survey.

Recently, our world witnessed major events that attracted a lot of attention towards the importance of automatic crowd scene analysis. For example, the COVID-19 breakout and public events require an automatic system to manage, count, secure, and track a crowd that shares the same area. However, analyzing crowd scenes is very challenging due to heavy occlusion, complex behaviors, and posture changes. This paper surveys deep learning-based methods for analyzing crowded scenes. The reviewed methods are categorized as (1) crowd counting and (2) crowd actions recognition. Moreover, crowd scene datasets are surveyed. In additional to the above surveys, this paper proposes an evaluation metric for crowd scene analysis methods. This metric estimates the difference between calculated crowed count and actual count in crowd scene videos.