Peijun Ye’s research while affiliated with Chinese Academy of Sciences and other places

The application of Unmanned Aerial Vehicles (UAVs) is crucial in traffic information collection. In addressing the challenge of detecting small targets in UAV imagery, simply increasing the model depth is not the most optimal solution. In this work, we propose MSDet, a novel object detection method based on master-slave backbone and bifurcation fusion. Different feature extraction methods provide varying feature information, and their fusion enables a more comprehensive and multidimensional description of the target. The simplified auxiliary networks are connected layer by layer with the main backbone, and their final output is fed back to the initial feature map. The main backbone and auxiliary networks can be flexibly selected and combined to adapt to the unique features of different scenarios. Bifurcation fusion achieves flexible multi-scale feature fusion by introducing branches during the top-down fusion process. One branch performs deeper top-down fusion to capture more shallow features, whereas the opposing branch offers a comprehensive understanding of the overall structure. Experimental results demonstrate that the proposed method outperforms state-of-the-art methods when applied to three UAV datasets. Furthermore, this study suggests that integrating with different backbones may yield better performance than simply scaling up models when faced with challenging situations.

The era has been witnessing our entering into a more professional and cooperative society and also a bloom of various digital humans that help us complete more challenging industrial work than ever. As a major human–machine communication channel, virtual–real interaction between digital and real humans is an efficient way to adjust the mismatch between specialized tasks and participants with unsuitable personalities, elevating the overall performance of such human participated systems. This article proposes a new paradigm for temporary personality shaping by prescriptively interacting the human operator/user with his “digital assistant” in cyber space. As a first proposed personality shaping with AI-aided approach, the new paradigm involves digital human modeling with uncertain personality, computational experiments on personality evolution, and prescriptive interaction with user counterpart. By iteratively and repeatedly executing the three steps, the personality of human participants is gradually tailored and prescribed so that he can well undertake specific tasks. Experiments based on online social media data and human–machine shared driving have indicated that our new personality shaping paradigm is feasible and effective in performing representative tasks.

Affected by people’s dynamic social activities, the imbalance between vehicle supply and demand in the Mobility-On-Demand(MOD) system is a common phenomenon. To improve traffic efficiency, an Auxiliary Network Enhanced Hierarchical Graph Reinforcement Learning (AHGRL) method is proposed for vehicle repositioning. Firstly, a hierarchical graph reinforcement learning (HGRL) framework is designed. The complex vehicle repositioning problem in real road networks is divided into many sub-tasks and multiple reinforcement learning algorithms are designed to solve decision problems of different levels. Traffic congestion is also considered and road nodes are clustered dynamically. And then an auxiliary graph reinforcement learning (AGRL) algorithm is designed for the actuator. It contains the prediction branch and the repositioning branch. States and rewards of agents could be designed accurately with the support of the prediction branch. The two branches cooperate in an auxiliary way to achieve excellent forecasting and repositioning effects. Finally, to enable efficient multi-vehicle coordination, a discrete Soft Actor-Critic algorithm is adopted in the repositioning branch, which learns multiple optimal actions for vehicles in the same area. Comparative experiments with real data demonstrate the effectiveness of our method. And ablation experiments verify the effectiveness and universality of the HGRL framework and the AGRL algorithm.

The coronavirus disease 2019 (COVID-19) pandemic has changed the world deeply. Urban trajectory big data collected by wireless sensing devices provide great assistance for COVID-19 prevention. However, except for contact tracing, trajectory data are rarely employed in other preventative scenarios against the pandemic. In this article, we try to extend the application of trajectories auto-collected by wireless sensing devices and simulate the epidemic spread in a trajectory data-driven manner. After that, the effects of three nonpharmacological measures are quantified. In contrast to existing studies, additional requirements such as the complex topological networks are needless in our simulation, where the interactions between agents are derived by the intersections of their trajectories. Concretely, the dynamic of virus propagation among individuals is first modeled, and then an agent-based microsimulation environment is built as an artificial system to conduct the epidemic spread simulation. Finally, the trajectories are loaded into the agents as the reliance for their interactions, and the macroscopic changes under different interventions are revealed in a bottom–up way. As a case study, we conduct the simulation based on the trajectories in a real region, in which we find the following. 1) Among the three examined nonpharmacological interventions, community containment is more effective than keeping social distance, which can lower the deaths to nearly 1/9 compared to no action, while travel restrictions play limited roles. 2) There is a strong positive correlation between population densities and mortality. 3) The timing of community containment triggered by confirmed diagnoses is proportional to the number of deaths, thus early containment will significantly decrease mortality.

With the breakthroughs in deep learning, recent years have witnessed a booming of artificial intelligence applications and services designed for vehicles. More recently, with the proliferation of vehicular networks, edge computing, and digital twin, billions of vehicular end devices (e.g., autonomous vehicles, roadside equipments) with high level of intelligence are interconnected and are promoted into the cyberspace or digital twin. Driving by this trend, there is an urgent need to push individual vehicular intelligence to collective vehicular intelligence so as to fully unleash the potential of AI. The resulting new interdiscipline, the internet of vehicular intelligence (IoVI), is beginning to receive a tremendous amount of interest. However, research on IoVI is still in its infancy stage, and the architecture and key features of IoVI are highly desired. To this end, we held a Distributed/Decentralized Hybrid Workshop on Intelligent Vehicles for Social Transportation (DHW-IVST), and this letter summarizes the outcomes of our discussion. Here, we present an overview of the IoVI architectures, frameworks, and virtual vehicle that counts for the features and characteristics of both driver and vehicle in the digital twin.

The interpretability of decision-making in autonomous driving is crucial for the building of virtual driver, promoting the trust worth of artificial intelligence (AI) and the efficiency of human-machine interaction. However, current data-driven methods such as deep reinforcement learning (DRL) directly acquire driving policies from collected data, where the decision-making process is vague for safety validation. To address this issue, this paper proposes cognitive reinforcement learning that can both simulate the human driver’s deliberation and provide interpretability of the virtual driver’s behaviors. The new method involves cognitive modeling, reinforcement learning and reasoning path extraction. Experiments on the virtual driving environment indicate that our method can semantically interpret the virtual driver’s behaviors. The results show that the proposed cognitive reinforcement learning model combines the interpretability of cognitive models with the learning capability of reinforcement learning, providing a new approach for the construction of trustworthy virtual drivers.

An intelligent human machine shared control system should be able to identify the driving style of real human drivers, so that the recommended driving strategy is in line with their driving style, and thus reduce the human intervention during the human-machine co-driving. The upgrade of vehicle sensors provides us an opportunity for a better judgement of the driver's driving style. This paper proposes a new driving style classification method based on the active learning. This method first uses the concept of driving volatility measurement to measure the performance of vehicles in terms of speed and acceleration, and then generates data to reduce the influence of data distribution on learning effect. The proposed method divides drivers' behaviors into aggressive, ordinary and calm styles. It can effectively learn the features related to driving style in driving volatility measurement, and the active learning method itself can reduce the requirement of human expert experience. The effectiveness of the proposed method is proved by the experiments on the UAH-Driveset, and then this paper uses the proposed method to analyze some data from the Security Pilot Model Deployment (SPMD).