Skeleton-based human action recognition has attracted increasing attention in recent years. However, most existing methods focus on single-person scenarios and struggle with complex behaviors in multi-person groups. In particular, they lack the capability to automatically identify and model core person. To address these challenges, this paper proposes a star-shaped group interaction model for skeleton-based action recognition. Firstly, the character importance scoring system analyzes both individual and group aspects: it evaluates each person's individual importance based on motion intensity and motion complexity, and assesses their significance within the group using centrality and interactivity. This process enables accurate identification of the core person in the video. Secondly, a core-star interaction graph is constructed with the core person as the center node and other individuals as peripheral nodes. The relationships among individuals are categorized into self-connections, centripetal connections, and centrifugal connections. For each type of connection, we design differentiated data augmentation strategies to fully exploit diverse action and interaction features. Finally, the structured skeleton data is fed into the star-shaped spatio-temporal graph convolutional network for efficient feature extraction and action classification. Experiments on several public benchmark datasets demonstrate that our method achieves state-of-the-art performance, achieving accuracies of 79.1%, 96.1%, and 93.1% on the NBA, Volleyball, and Volleyball-weak datasets, respectively.

Currently, fine-grained skeleton action recognition based on graph convolutional networks (GCNs) has become an important research focus. Fine-grained action recognition refers to the accurate recognition of subtle, complex or detailed actions. This task is particularly challenging due to the limited appearance information in skeleton data and the limitations of predefined single-topology skeleton structures. To address these challenges, we propose an action-responsive contrastive network (ARCN). The network consists of two main components: an action-responsive graph convolutional network (ARGCN) with enhanced skeleton topology and a fine-grained action comparator (FAC) that uses feature contrastive learning to explore the latent space of motion features. The ARGCN contains two specialized modules: the action-responsive topology (ART) module, which captures important motion features through the learned action-specific topology structure matrix and multiscale temporal features; and the action-responsive attention (ARA) module, which learns complex spatiotemporal skeleton attention information. These modules jointly generate a multichannel cross-temporal dynamic skeleton joint attention topology map tailored for the specific action being analysed. To further clarify the fine-grained action feature differences, the FAC is integrated in some stages of the ARGCN. The FAC performs spatiotemporal decoupling of feature maps, classifies and contrasts similar and different fine-grained motion features, and builds a learnable latent space for fine-grained motion, thereby improving classification performance. Our model is evaluated on six public datasets: NTU RGB+D, NTU RGB+D 120, NW-UCLA, UAV-Human, Finegym, and Diving48. It achieves 91.2% accuracy on the NTU RGB+D 120 dataset X-Set, 97.2% accuracy on the NW-UCLA dataset, 44.6% accuracy on the UAV-Human dataset CSv1, 72.0% accuracy on the UAV-Human dataset CSv2, 95.3% accuracy on the Finegym dataset, and 54.3% accuracy on the Diving48 dataset, which are competitive results compared with the state-of-the-art methods.

Like deep neural network (DNN)-based classifiers, DNN-based trackers are also vulnerable to adversarial attacks that degrade the tracking performance by adding adversarial perturbations to the input videos. This paper proposes a detection method for the first time to assist the tracker in detecting adversarial attacks. The adversarial perturbations in the visual object tracking task are invisible but are effective at attacking trackers. This naturally creates challenges in detecting attacks in the spatial pixel domain. To this end, we innovatively transfer the detection of adversarial attacks from the spatial domain to the frequency domain. Specifically, we first theoretically prove that the perturbations are added mainly to the high-frequency band of the video. Then, from the empirical studies, we conclude that the low-frequency band contributes most to the tracking performance and is most robust against adversarial attacks. According to the theoretical proof and empirical conclusion, we finally design an unsupervised adversarial detection framework, which mainly contains a frequency decomposition module (FDM), a target tracker (TT) with its mirror tracker (MT), and a discriminant module (DM). For an input video, the TT is fed the full-frequency video, whereas the MT takes as input the low-frequency video that is decomposed by the FDM. The DM discriminates the input video as adversarial or natural by comparing the racking performance differences between the two trackers. The whole detection process is performed along with the tracking phase, and all the modules in the framework require no training on adversarial examples. Extensive experiments demonstrate that our adversarial detection framework can effectively detect mainstream adversarial attacks in the tracking field. It can also be flexibly integrated with many trackers, including anchor-based and anchor-free trackers. More importantly, the trackers integrated with the detection framework can still maintain near-original tracking performance.

As a challenging visual task, visual object tracking has recently been composed of the classification and regression subtasks. The anchor-free regression network gets rid of the dependence on the anchors, but the redundant range makes it usually regress some samples involving non-target information. Evenly dividing a target by the regular receptive field often causes ambiguous target localization. To address these issues, we propose a regression-selective feature-adaptive tracker (RSFA), where the regression-selective subnetwork can not only free the regression task from anchors, but can also select more effective regression samples using the refined criterion. The proposed feature-adaptive strategy concentrates the classification subnetwork on target feature extraction via adaptively modifying the receptive field, and the attached centrality branch offers a correction for target localization by exploiting the spatial information. Additionally, the designed online update mechanism realizes the tracker's online optimization, improving robustness against target deformation. Extensive experiments are conducted on challenging benchmarks, including GOT10 K, OTB2015, UAV123, NFS, VOT2018, VOT2019 and VOT2020-ST. Our tracker achieves satisfactory tracking results, and the evaluations of its tracking performance rank first or second in comparison with the state-of-the-art tracking algorithms.