According to historical crowd accidents data, pedestrian panic behaviors have often triggered and further exacerbated crowd incidents, leading to severe casualties in transit hubs. To date, most existing panic behavior recognition models still fail to detect crowd panic quickly enough to guide timely pedestrian-control measures. To address this gap, we propose the pedestrian Panic Behavior Recognition Model (PBRM), which integrates audio-semantic signals, speech-derived panic keywords, 21-keypoint kinematic patterns associated with running and falls, and multi-inflection pedestrian trajectories. These heterogeneous cues are consolidated through a lightweight fusion mechanism—efficient enough for real-time execution on edge devices—to generate a unified panic score, thereby providing operators with a rapid and reliable basis for crowd-control decisions. Evaluated on the four-hour Hongqiao-Panic dataset collected at Hongqiao Railway Station, PBRM achieves 84.2% precision, 85.7% recall, 84.9% F1, and 1.18 s average alert lead, improving the best single-modality baseline by 3.0 pp F1 while keeping the false-alarm rate (FAR) below 8%. Ablation study shows that audio-semantic channels dominate early cues, where as posture-trajectory streams reinforce decision robustness under occlusion. These results demonstrate that PBRM can be deployed on edge devices to provide operators with a critical lead time for crowd-control interventions, advancing intelligent transportation safety from post-event surveillance to proactive prevention.

Panic behaviors in a pedestrian flow often lead to a state of chaos or disorder among the pedestrian crowd, resulting in a crowd accident with high possibility. To investigate the panic pedestrian dynamics and further prevent serious crowd accidents, simulation based on dynamics modeling and accident video data is a popular solution to date. Thereby, it is challenging but significant to improve the crowd dynamics model more consistent with the ground truth of real pedestrian movement scenarios, with consideration of both physical contact force and panic emotional propagation in a crowd. Therefore, this study proposed an extended social force model (ESFM) by applying the physical contact-force estimation during pedestrian collision based on non-smooth contact dynamics. Subsequently, the ESFM was integrated with an improved panic propagation model (IPPM) considering obstacle and promotion factors. Finally, taking the crowd panic accident happened in Nepal in 2015 as an experiment case, the simulation of panic crowd dynamics was conducted within Anylogic software. Four cases of SFM, ESFM, SFM+IPPM, and ESFM+IPPM were compared quantitatively and graphically. The experimental results showed that the pedestrian distribution obtained from the proposed ESFM+IPPM was the closest to the ground truth during the panic response period, with 28.8% lower of Hausdorff distance than the original SFM, and 21.6% lower the well-known BHSFM, respectively. This approach can help improve the panic crowd modeling and pedestrian distribution prediction in real scenarios.

A fall is an abnormal behavior that rarely occurs, however, once it occurs in a crowded scenario, it is prone to cause local disturbance, density-velocity fluctuations, and crowd instability even leading to a stampede accident. Currently, research on fall behavior mainly depends on single-point detection approach but lacks investigation on disturbing mechanism in a crowd flow. To address this issue, this paper first proposed a pressure dynamics model based on limb-level contact to analyze pedestrian behavioral characteristics. Then, based on the random Brownian motion, the disturbance distribution of fall behavior was determined theoretically. Using two parameters (a pressure coefficient and disturbance intensity), the pressure term of the fluid dynamics Aw-Rascle model was improved, and the internal disturbance propagation model (DPM) of fall behavior was established, as a primary contribution of this study. Further, to eliminate the disturbance propagation in a crowd, damping motion theory was employed. The disturbance elimination mechanism of abnormal behavior was discussed to suppress disturbance propagation caused by a pedestrian fall behavior. To validate the proposed model, both field experiment and numerical simulation were conducted respectively. The stability performance of the proposed DPM was evaluated with standard deviation (less than 0.1631) based on 50 repetitive experiments. Results showed that this study could help discover the disturbance propagation dynamics and elimination mechanisms of pedestrian fall behavior in a crowded scenario.

With the increasing number of video surveillance cameras in public buildings, it has become challenging, yet significant to detect abnormal pedestrian behaviors in crowd management, to prevent crowd accidents. Although current advancements in human action recognition based on computer vision can help detect abnormal behaviors after their incidence, majority of them lack the ability to detect potential characteristics prior to the occurrence of real abnormal behaviors. Hence, in this study, we addressed this issue by proposing a novel dynamic centroid model (DCM) of a human body, and rebuilding pedestrian joint sub-segments from human skeleton key nodes obtained in camera images. We built a weighted centroid-combined force model based on Newton’s second law, considering acceleration, mass inertial of human body sub-segments, and internal constraints. Thereafter, pedestrian kinematic and dynamic parameters were analyzed, such as speed, trajectory, force. Furthermore, abnormal behavior detection criteria were constructed for typical abnormal-behavior cases: U-turn and fall-down. Comparative experiments between the proposed DCM and the state-of-the-art methods were conducted. The experimental results showed that the model was capable of detecting abnormal behaviors, with mean values of lead time of 277 ms in U-turn behavior, and 562 ms in fall-down behavior, prior to the captured occurrence of these two abnormal behaviors. Finally, a de-occlusion algorithm was designed and jointly used with DCM, validated by a fall-down detecting experiment including partial occlusion. Therefore, this study holds significant value for the prevention of abnormal pedestrian behaviors in public places.

With the increase in inevitable large-scale crowd aggregation, disastrous pedestrian stampedes occurred with increasing frequency over the past decade. To prevent these tragedies, it is significant to assess crowd accident-risk (CAR) and identify high-risk areas to control crowd flow dynamically. The cost function of a conventional fluid dynamics model is improved with new items of Gaussian white noise and protection factor, considering both the abnormal pedestrian movements and social distance control due to epidemic, thereby to establish an improved crowd flow model comprehensively. Different from conventional density-based pedestrian aggregation-risk models, this study proposes a hybrid crowd accident-risk assessment (HCRA) model based on internal energy and information entropy. Using the HCRA model, we can consider not only crowd density but also the modulus and direction of a crowd velocity vector simultaneously. Then this study designs a framework to realize crowd accident risk assessment based on the improved crowd-flow model and HCRA model. To validate the proposed models, case studies of CAR assessment in the large-scale waiting hall of the Shanghai Hongqiao railway station are conducted. The pedestrian social control distance-range of 1.0 m–2.0 m under the COVID-19 epidemic situation is verified numerically. Moreover, a valuable result is that this social control distance-range can be shortened to 1.0 m–1.9 m without increase of crow accident-risk. Subsequently, the down-limit of accommodation-capacity of this large waiting hall can be enhanced to 10.54% under this epidemic.

This study investigates a dynamic flow model for crowd evacuation at T-shaped street junctions (TSJs) from a macroscopic view. The Aw-Rascle traffic flow model is modified by constructing an impact matrix in the street intersection area to practically describe the crowd convergence mechanism at a TSJ. For coherence, this modified model is proved to be anisotropic, similar to the original Aw-Rascle traffic flow model. To describe real scenarios with higher crowd density and lower speeds during organized pilgrimages, the initial Gaussian distribution of the crowd is improved to a higher-order smoothing function. To validate the modified Aw-Rascle traffic flow model, we reconstruct the drastic stampede that occurred at the TSJ of streets 204 and 223 during the 2015 Mecca pilgrimage. Further, the main environmental parameters that potentially lead to a stampede are discussed with numerical simulations. A valuable suggestion is that the street width ratio should be extended from 1.1 to 1.4 to prevent stampedes, matching the expansion engineering of street 204 reported by BBC News. An interesting phenomenon is that the closer the bus unloading location on street 223 is to the TSJ center, the lower the maximum crowd density and the safer the pedestrians will be. With this modified Aw-Rascle flow model at TSJs, this paper provides strategic and technical suggestions for future crowd flow control to reduce the risk of crowd stampedes.

Network function virtualization integrates different types of dedicated network equipment into standard industry IT server, storage and switch equipment, enabling network functions traditionally implemented using specific equipment to use software running on IT industry standard server hardware, thereby enhancing system flexibility. Using the organic combination of NFV and software-defined network technologies, an software defined Smart grid communication network can be constructed, so that the network functions of service function chaining can be implemented on general-purpose equipment, and end-to-end services are transformed into a set of sequentially connected VNFs, which can effectively deploy and manage service function chains. Service provision and server resource utilization will be affected by SFC mapping. In order to ensure the reasonable use of network resources and the QoS of SFC, the research on SFC mapping algorithms is particularly important. In this paper, we propose a service chain mapping algorithm based on reinforcement learning, which aims to learn by the system status and the feedback value given by the mapped environment and then finally determine the actual deployment location of each virtual function node in the SFC. The comparison and analysis with other algorithms show that the SFC mapping algorithm proposed in this paper can adjust the feedback value function to optimize the load balance of the system and reduce the SFC average link delay in different network topologies.

Recent successes of deep learning-based recognition rely on maintaining the content related to the main-task label. However, how to explicitly dispel the noisy signals for better generalization remains an open issue. We systematically summarize the detrimental factors as task-relevant/irrelevant semantic variations and unspecified latent variation. In this paper, we cast these problems as an adversarial minimax game in the latent space. Specifically, we propose equipping an end-to-end conditional adversarial network with the ability to decompose an input sample into three complementary parts. The discriminative representation inherits the desired invariance property guided by prior knowledge of the task, which is marginally independent to the task-relevant/irrelevant semantic and latent variations. Our proposed framework achieves top performance on a serial of tasks, including digits recognition, lighting, makeup, disguise-tolerant face recognition, and facial attributes recognition.

We consider the problem of comparing the similarity of image sets with variable-quantity, quality and un-ordered heterogeneous images. We use feature restructuring to exploit the correlations of both inner&inter-set images. Specifically, the residual self-attention can effectively restructure the features using the other features within a set to emphasize the discriminative images and eliminate the redundancy. Then, a sparse/collaborative learning-based dependency-guided representation scheme reconstructs the probe features conditional to the gallery features in order to adaptively align the two sets. This enables our framework to be compatible with both verification and open-set identification. We show that the parametric self-attention network and non-parametric dictionary learning can be trained end-to-end by a unified alternative optimization scheme, and that the full framework is permutation-invariant. In the numerical experiments we conducted, our method achieves top performance on competitive image set/video-based face recognition and person re-identification benchmarks.

This paper targets the task with discrete and periodic class labels (e.g., pose/orientation estimation) in the context of deep learning. The commonly used cross-entropy or regression loss is not well matched to this problem as they ignore the periodic nature of the labels and the class similarity, or assume labels are continuous value. We propose to incorporate inter-class correlations in a Wasserstein training framework by pre-defining (i.e., using arc length of a circle) or adaptively learning the ground metric. We extend the ground metric as a linear, convex or concave increasing function w.r.t. arc length from an optimization perspective. We also propose to construct the conservative target labels which model the inlier and outlier noises using a wrapped unimodal-uniform mixture distribution. Unlike the one-hot setting, the conservative label makes the computation of Wasserstein distance more challenging. We systematically conclude the practical closed-form solution of Wasserstein distance for pose data with either one-hot or conservative target label. We evaluate our method on head, body, vehicle and 3D object pose benchmarks with exhaustive ablation studies. The Wasserstein loss obtaining superior performance over the current methods, especially using convex mapping function for ground metric, conservative label, and closed-form solution.

Various representation-based methods have been developed and shown great potential for pattern classification. To further improve their discriminability, we propose a Bi-level optimization framework in terms of both low-dimensional projection and collaborative representation. Specifically, during the projection phase, we try to minimize the intra-class similarity and inter-class dissimilarity, while in the representation phase, our goal is to achieve the lowest correlation of the representation results. Solving this joint optimization mutually reinforces both aspects of feature projection and representation. Experiments on face recognition, object categorization and scene classification dataset demonstrate remarkable performance improvements led by the proposed framework.

Effective training of the deep neural networks requires much data to avoid underdetermined and poor generalization. Data Augmentation alleviates this by using existing data more effectively. However standard data augmentation produces only limited plausible alternative data by for example, flipping, distorting, adding noise to, cropping a patch from the original samples. In this paper, we introduce the adversarial autoencoder (AAE) to impose the feature representations with uniform distribution and apply the linear interpolation on latent space, which is potential to generate a much broader set of augmentations for image classification. As a possible “recognition via generation” framework, it has potentials for several other classification tasks. Our experiments on the ILSVRC 2012, CIFAR-10 datasets show that the latent space interpolation (LSI) improves the generalization and performance of state-of-the-art deep neural networks.

To date there is no imaging modality for cardiac arrhythmias which remain the leading cause of sudden death in the United States (> 300000/yr.). Electrocardiographic imaging (ECGI), a noninvasive modality that images cardiac arrhythmias from body surface potentials, requires the geometrical relationship between the heart surface and the positions of body surface ECG electrodes. A photographic method was validated in a mannequin and used to determine the three-dimensional coordinates of body surface ECG electrodes to within 1 mm of their actual positions. Since fluoroscopy is available in the cardiac electrophysiology (EP) laboratory where diagnosis and treatment of cardiac arrhythmias is conducted, a fluoroscopic method to determine the heart surface geometry was developed based on projective geometry, epipolar geometry, point reconstruction, b-spline interpolation and visualization. Fluoroscopy-reconstructed hearts in a phantom and a human subject were validated using high-resolution computed tomography (CT) imaging. The mean absolute distance error for the fluoroscopy-reconstructed heart relative to the CT heart was 4 mm (phantom) and 10 mm (human). In the human, ECGI images of normal cardiac electrical activity on the fluoroscopy-reconstructed heart showed close correlation with those obtained on the CT heart. Results demonstrate the feasibility of this approach for clinical noninvasive imaging of cardiac arrhythmias in the interventional EP laboratory.