In the field of skeleton-based gesture recognition, occlusion remains a significant challenge, significantly degrading performance when key joints are occluded or disturbed. To tackle this issue, we propose DiffTrans, a practical conditional diffusion model for occlusion recognition, which enables skeleton-based gesture recognition under high occlusion by generating more likely samples. This study addresses the hand skeleton occlusion problem by framing it as a conditional denoising problem, where unoccluded data serve as observations and occluded data as repair targets. We employ a conditional diffusion model to impute the missing skeleton data and the DSTANet model, which is based on the transformer, to learn the skeleton feature representations. Research results show that the DiffTrans outperforms existing methods under various occlusion modes, maintaining high performance even in scenarios with a high missing rate.

This study focuses on trajectory planning and motion control policies in autonomous racing, which necessitates pushing the capacity boundaries of racing vehicles to achieve maximum speeds and minimal lap times. We propose an innovative planning control framework that integrates risk-conscious mutations in jump-start reinforcement learning (RCM-JSRL) and nonlinear model predictive control (NMPC). The RCM-JSRL algorithm incorporates jump-start curriculum learning and the risk-conscious genetic algorithm into reinforcement learning, leveraging prior expert knowledge and a curiosity-driven exploration mechanism to enhance training efficiency while avoiding excessively conservative policy generation in high-complexity and high-risk scenarios. NMPC generates locally optimal control commands that adhere to vehicle dynamics constraints while following the designated trajectory. Following training on track maps with varying difficulty levels, the proposed controller successfully executes a superior policy compared to the guide policy, providing evidence of its effectiveness and scalability. It is our belief that this technology can be applied in everyday driving scenarios, improving efficiency under special conditions, ensuring stability in critical situations, and broadening the scope of autonomous driving applications.

This study introduces the Metacognitive-Attribution Inspired Reinforcement Learning (MAIRL) approach, designed to address unprotected interactive left turns at intersections—one of the most challenging tasks in autonomous driving. By integrating the Metacognitive Theory and Attribution Theory from the psychology field with reinforcement learning, this study enriches the learning mechanisms of autonomous vehicles with human cognitive processes. Specifically, it applies Metacognitive Theory’s three core elements—Metacognitive Knowledge, Metacognitive Monitoring, and Metacognitive Reflection—to enhance the control framework’s capabilities in skill differentiation, real-time assessment, and adaptive learning for interactive motion planning. Furthermore, inspired by Attribution Theory, it decomposes the reward system in RL algorithms into three components: 1) skill improvement, 2) existing ability, and 3) environmental stochasticity. This framework emulates human learning and behavior adjustment, incorporating a deeper cognitive emulation into reinforcement algorithms to foster a unified cognitive structure and control strategy. Contrastive tests conducted in various intersection scenarios with differing traffic densities demonstrated the superior performance of the proposed controller, which outperformed baseline algorithms in success rates and had lower collision and timeout incidents. This interdisciplinary approach not only enhances the understanding and applicability of RL algorithms but also represents a meaningful step towards modeling advanced human cognitive processes in the field of autonomous driving.

Temporal action detection, a critical task in video activity understanding, is typically divided into two stages: proposal generation and classification. However, most existing methods overlook the importance of information transfer among proposals during classification, often treating each proposal in isolation, which hampers accurate label prediction. In this article, we propose a novel method for inferring semantic relationships both within and between action proposals, guiding the fusion of action proposal features accordingly. Building on this approach, we introduce the Proposal Semantic Relationship Graph Network (PSRGN), an end-to-end model that leverages intra-proposal semantic relationship graphs to extract cross-scale temporal context and an inter-proposal semantic relationship graph to incorporate complementary neighboring information, significantly improving proposal feature quality and overall detection performance. This is the first method to apply graph structure learning in temporal action detection, adaptively constructing the inter-proposal semantic graph. Extensive experiments on two datasets demonstrate the effectiveness of our approach, achieving state-of-the-art (SOTA). Code and results are available at http://github.com/Riiick2011/PSRGN .

This study focuses on the motion planning and risk evaluation of unprotected left turns at occluded intersections for autonomous vehicles. In this paper, we present an interactive motion planning controller that combines Cross-Observability Optimistic-Pessimistic Safe Reinforcement Learning (COOP-SRL) and Nonlinear Model Predictive Control (NMPC), with consideration of the uncertain potential risk of occluded zone, the trade-off between safety and efficiency, and the dynamic interaction between vehicles. The proposed COOP-SRL algorithm integrates fully and partially observable policies through cross-observability soft imitation learning to leverage the expert guidance and improve learning efficiency. Moreover, the optimistic exploration policy and pessimism safe constraint are adopted to provide an adaptive safe strategy without hindering the exploration during learning process. Finally, the evaluations of the proposed controller were conducted in occluded intersection scenarios with various traffic density level, which indicate that the proposed method outperforms both the optimization-based and learning-based baselines in qualitative and quantitative indexes.

Temporal action detection is a challenging task in video understanding, which is usually divided into two stages: proposal generation and classification. Learning proposal features is a crucial step for both stages. However, most methods ignore temporal information of proposals and consider background and action frames in proposals equally, leading to poor proposal features. In this paper, we propose a novel Temporal Attention-Pyramid Pooling (TAPP) method to learn proposal features of arbitrary length action proposals. The TAPP method exploits the attention mechanism to focus on the discriminative part of proposals, suppressing background influence on proposal features. It constructs a temporal pyramid structure to convert arbitrary length proposal feature sequences to multiple fixed-length sequences while retaining the temporal information. In the TAPP method, we design a multi-scale temporal function and apply it to the temporal pyramid to generate final proposal features. Based on the TAPP method, we construct a temporal action proposal generation model and an action proposal classification model, and then we perform extensive experiments on two mainstream temporal action detection datasets for the temporal action proposal and temporal action detection tasks to verify our models. On the THUMOS’14 dataset, our models based on the TAPP significantly outperform the previous state-of-the-art methods for both tasks.