Collaborative perception has garnered significant attention as a crucial technology to overcome the perceptual limitations of single-agent systems. Many state-of-the-art (SOTA) methods have achieved communication efficiency and high performance via intermediate fusion. However, they share a critical vulnerability: their performance degrades under adverse communication conditions due to the misalignment induced by data transmission, which severely hampers their practical deployment. To bridge this gap, we re-examine different fusion paradigms, and recover that the strengths of intermediate and late fusion are not a trade-off, but a complementary pairing. Based on this key insight, we propose CoRA, a novel collaborative robust architecture with a hybrid approach to decouple performance from robustness with low communication. It is composed of two components: a feature-level fusion branch and an object-level correction branch. Its first branch selects critical features and fuses them efficiently to ensure both performance and scalability. The second branch leverages semantic relevance to correct spatial displacements, guaranteeing resilience against pose errors. Experiments demonstrate the superiority of CoRA. Under extreme scenarios, CoRA improves upon its baseline performance by approximately 19% in [email protected] with more than 5x less communication volume, which makes it a promising solution for robust collaborative perception.

Millimeter-wave (mmWave) 5G-V2X communications face significant challenges in real-time beam alignment within high-mobility vehicular networks. While environmentaware beam prediction methods mitigate channel estimation overhead, their efficacy is severely compromised by modality data loss stemming from lighting variations, adverse weather, or sensor failures. To address this issue, we propose a Cross-modal Losstolerant Beam Prediction model (CLBP). CLBP robustly fuses RGB camera and LiDAR data, employing a novel cross-modal attention mechanism to achieve resilient feature alignment across these heterogeneous modalities. Furthermore, a Branch Features Dynamic Fusion (BFDF) module adaptively reweights modality features, suppressing noise from degraded inputs and promoting effective information propagation to enhance resilience. To facilitate realistic evaluation, we introduce a Data-Conditioned Missingness Mechanism (DCMM), which augments the DeepSense 6G V2V dataset with meticulously simulated sensor failure scenarios. Experimental results demonstrate CLBP's superior performance, achieving 94.48% Top-5 beam prediction accuracy even under 10% modality loss, and a 29% reduction in average power loss compared to baseline methods. These findings demonstrate CLBP's significant robustness in dynamic vehicular environments and its capacity to maintain consistent, high-performance beam prediction despite challenging data imperfections.

Cooperative perception facilitated by vehicle-to-vehicle (V2V) data sharing has emerged as a crucial enabler for safe and efficient autonomous driving. However, the current state-of-the-art algorithms are unable to resolve severe performance degradation caused by communication impairments in realistic V2V scenarios. This paper models the V2V communication quality and confirms their fragile robustness under loss conditions. To this end, we propose a robust and efficient cooperative perception framework. Specifically, we propose RoCooper, a robust fusion algorithm. It leverages the lossless ego feature as an anchoring foundation, then utilizes multi-dimensional feature correlations and dynamic regional selective cross-learning. This allows it to perform multi-scale feature recovery and judiciously fuse multi-view features from neighboring vehicles. In addition, we design ReduAdapt, an efficient redundancy-aware scheduling scheme that first constructs hierarchical metadata to decompose 3D perception space, then applies dynamic multi-factor thresholds for region-specific cropping, and finally performs retention-driven adaptive compression, enabling connect vehicles to prioritize critical data streams while dynamically suppressing redundant transmissions, thereby maximizing effective throughput. Extensive evaluations of real-world datasets demonstrate that our method achieves state-of-the-art performance in varying impairment scenarios, while delivering more efficient compression-transmission at comparable levels.

Visual perception, as a core component of Intelligent Transportation Systems (ITS), plays a key role in enhancing safety and efficiency in urban mobility. While single-task visual perception methods have applications in areas like pedestrian detection and traffic sign recognition, the complexity of real-world scenarios necessitates a shift toward multi-task approaches. This paper introduces the Edge-assisted Multi-task Visual Perception (EMVP) system, which is specifically designed to address the computational intensity and dynamic concurrency challenges inherent to multi-task processing in edge environments. EMVP adopts a collaborative architecture that strategically partitions computational tasks between vehicles and Road-Side Units (RSUs). By integrating Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), the system achieves a lightweight yet efficient multi-task model for resource-constrained environments. To adapt to the dynamic and concurrent nature of multi-vehicle scenarios, EMVP incorporates a content-aware adaptive inference mechanism based on reinforcement learning, enabling dynamic task scheduling to improve Quality of Service (QoS). Experimental results demonstrate that, compared to the single-task baseline model, the multi-task model of EMVP reduces the computational cost by 86.59% on average while achieving a 4.32% improvement in accuracy. Additionally, in dynamic multi-access environments, EMVP’s adaptive scheduling mechanism, which leverages spatiotemporal content awareness, achieves an average QoS improvement of 7.36% over the sub-optimal method.

Adaptive Bitrate (ABR) is an effective way to improve the user’s viewing experience. In recent years, with the increasing demand for high quality and maturing audio-video technology, video-oriented ABR algorithms focusing on various scenarios have been widely studied. However, in commuting scenarios, users’ internal demand for audio-video optimization and the external factor of network fluctuations make it difficult for traditional video-oriented ABR algorithms to cope. To address these scenarios’ problems, we propose a collaborative audio-video ABR strategy under commuting (COAV). It includes two agents deciding audio and video bitrates respectively to reply to the internal demand and an embeddable prediction module to address the external factor. COAV adopts a centralized training and decentralized execution framework, designs two actors and one critic reinforcement learning neural networks based on Advantage Actor-Critic algorithm, and uses TPA-LSTM to predict the throughput as neural networks state, so as to solve the problem of internal demand and external factor in depth. We evaluate the performance of COAV using trace data collected in commuting scenarios, and the experimental results show that COAV improves the average QoE by about 15.5% compared to RAV baseline.

The rapid development of real-time communications (RTC) has created many challenges for designing a proper transport control module, which determines how much media data can be sent in real time. Reinforcement learning (RL) -based transport control algorithms have shown great potential, but still face some unique challenges. For example, accurate bandwidth prediction is often necessary but it is difficult to guarantee accuracy due to bandwidth non-stationarity. In addition, how to alleviate the cold-start and overestimation problems of learning-based algorithms to achieve higher training efficiency is also a headache for researchers. In this work, we propose a new training framework that leverages the advanced Transformer model to capture the non-stationarity of the bandwidth sequence and improve the bandwidth prediction accuracy, while using knowledge distillation and transfer learning techniques to train the RL model efficiently and alleviate the cold-start problem of the model in the training environment. Besides, we employ the Double-Q learning mechanism to suppress the overestimation problem and further enhance the training efficiency. Based on this framework, we have trained a new RTC transport control algorithm NSAC and test it on our own platform. The experiments prove that NSAC adapts better to the unstable network environment than the state-of-the-art solutions. In conditions of weak network, the video throughput experiences a 14.11% increase, accompanied by reductions of 5.64%, 28.12%, and 25.86% in delay, loss rate, and stall rate, respectively. These improvements notably enhance the quality of user experience.

With the rise of artificial intelligence, various video analytics models have been applied in many fields. Numerous studies are preoccupied with expanding the size of the model to achieve greater accuracy, yet inference latency is unbearable when models are deployed to resource-constrained terminal devices. Edge computing ensures efficient and accurate inference by offloading video inference tasks to edge servers because of low network latency and high-performance hardware. However, edge-assisted video analytics systems encounter challenges due to the dynamic nature of video frames, fluctuating network signals, and the mismatch between arithmetic power and model size. To overcome these obstacles, we propose MDRL, an edge-assisted video analytics framework based on Multi-modal Deep Reinforcement Learning. MDRL adaptively determines the offloading strategy of video frames by observing multi-modal information from video frames and network signals and updates the parameters using Deep Reinforcement Learning (DRL) algorithms. We compare MDRL with various baselines and the experimental results show that MDRL has the highest overall optimization of latency, accuracy, and network bandwidth consumption in various experimental scenes.

Multi-user Augmented Reality (AR) applications allow multiple users to interact within the same physical environment. However, existing multi-user systems lack the ability to recognize the complex physical environment. It is also challenging to achieve real-time multi-user AR under dynamic network conditions and different computing workloads. In this paper, we present EMAR, an edge-assisted multi-user mobile AR system. The system uses an edge cooperative optimization strategy to improve the accuracy of scene recognition, and adopts adaptive offloading and AR dynamic update strategy to ensure the real-time and effectiveness of AR recovery. We implement EMAR on mobile smartphones and an edge server. We verify its performance in different environments. The results show that EMAR achieves precision above 84.7% and recall above 92% for accurately identifying the same scene. Compared to the baseline method, our system is more robust in the dynamic environment, showing significant performance improvements.

Highly accurate perception is one of the pivotal factors for the safe operation of Intelligent Connected Vehicles (ICVs). Nevertheless, occlusion blind spots, limited fields-of-view, and low-point density of the sensor data lead to limited perception for the single ICV, which can be well addressed with vehicle-to-vehicle (V2V) cooperative perception. Recent development of V2V cooperative perception technology have made the perception of ICVs more and more accurate and reliable. In V2V cooperative perception, LiDAR and Camera are two types of complementary sensors for ICVs. However, using only specific single-modal data such as Camera RGB images or LiDAR point clouds for V2V collaborative perception cannot fully improve perception accuracy. Furthermore, the large model based on Transformers has been proven to effectively enhance multi-modal fusion. To this end, we propose MCoT, a novel approach for multi-modal V2V cooperative perception with Transformers. Our MCoT extracts intermediate features from RGB images and point cloud of different agents, aligning them in the Bird’s-Eye View (BEV) perspective through rigid association. Subsequently, we use the cross-attention mechanism to perform a soft fusion of these features in the BEV domain. The attention mechanism empowers our model with the ability to adaptively discern which regions of the image and LiDAR are most relevant, and what information should be extracted from them. Extensive evaluations demonstrate that MCoT can significantly enhance the accuracy and robustness of perception. Our model achieved remarkable results on the large-scale simulation dataset OPV2V, improving the average accuracy by 71.43% compared to the baseline and outperforming the second-place by nearly 3.95%. Our approach also demonstrates the fastest convergence rate under the same number of training epochs.

On-device AI is taking over our daily lives by moving closer to mobile devices as perception applications. A data stream perception application generally has three essential requirements: timeliness, smoothness, and orderliness. Most researchers’ efforts to date have proposed various offloading approaches to accelerate compute-intensive AI algorithms in perception applications, thereby fulfilling the requirement of timeliness. However, the lack of concern about the smoothness and orderliness of the data stream will result in fluctuation and commotion anomalies that greatly impair the user experience. In this paper, we propose an enhanced Offloading System with sMoothness and Orderliness (OSMO) to guarantee perception applications’ smooth refresh rates while processing data streams in proper orders with low overhead. OSMO takes advantage of heterogeneous computing devices and data-level parallelism in the offloading process. A scheduling strategy is further devised that dynamically tunes a set of parameters to achieve the best trade-offs among the three requirements of perception applications. We implement a prototype system based on TensorFlow and its typical Android demos. Real-world evaluations demonstrate that our solution can effectively address the fluctuation and commotion issues while providing a high data processing rate with multi-device collaboration.

In recent years, Real-Time Communication (RTC) has been widely used in many scenarios, and Congestion Control (CC) is one of the important ways to improve the experience of such applications. Accurate bandwidth prediction is the key to CC schemes. However, designing an efficient congestion control scheme with accurate bandwidth prediction is challenging, largely because it is essentially a Partially Observable MDP (POMDP) problem, making it difficult to use traditional hand-crafted methods to solve. We propose a novel hybrid CC scheme LRCC, which combines attention-based Long Short-Term Memory (LSTM) and Reinforcement Learning (RL), realizing more accurate bandwidth prediction and congestion control by adding bandwidth memory information provided by the recurrent neural network to the RL decision-making process. Trace-driven experiments show that our proposed method can significantly reduce packet loss and improve bandwidth utilization in various network scenarios, outperforming baseline methods on overall QoE.

As a proper supplement to traditional wireless communication, opportunistic network provides a feasible and inexpensive way to achieve message delivery, especially in extreme environments. However, the gossip dissemination problem severely influences the network performance, and is hardly tackled due to the network characteristics of more transmission delay and higher mobility. To address this problem, we propose a robust and efficient framework named SCON, which contains a flexible region-based cluster routing algorithm to relieve the gossip impacts, crowd-sourcing prosecution and attacker judgment schemes and node reward and punishment mechanisms to discover and eliminate attackers that disseminate gossips, as well as several buffer maintenance mechanisms to further improve the network performance. Comprehensive evaluations demonstrate the high performance and robustness of our frame-work compared with the state-of-the-art approaches when gossip dissemination occurs.

Maximizing user quality of experience (QoE) is the ultimate goal of video players, and adaptive bitrate (ABR) is recognized as one of the most effective solutions. Approaches employing reinforcement learning (RL) have performed well as hybrid ABR algorithms, due to the ability to learn autonomously. However, throughput, which plays a crucial role in low-delay video streaming, is difficult to predict simply in mobile and wireless networks, and the inaccurately predicted throughput can lead to the wrong selection of bitrates. Worse, the general RL approaches are prone to frequent bitrate switching due to bandwidth fluctuation. These obstacles make the RL-based ABR approach unable to truly reflect the user QoE. We propose TP-RL, an application that makes ongoing decisions to maximize user QoE. To realize this, TP-RL adopts three ideas: (i) It takes the RL neural network as the main body of decision-making, which will inherit the advantages of RL and improve on this basis; (ii) Explore Mogrifier LSTM for throughput prediction, and replace the throughput part in the state space of the original RL neural network with a prediction module; (iii) The decided bitrate is further processed to achieve better smoothness when the bandwidth fluctuates. The performance of TP-RL is evaluated in different experimental environments, and experiments show that it can improve QoE by about 14% to 20.7% compared with the best baseline.

Indoor positioning technology based on Wi-Fi fingerprint recognition has been widely studied owing to the pervasiveness of hardware facilities and the ease of implementation of software technology. However, the similarity-based method is not sufficiently accurate, whereas the offline training of the neural network-based method is overly time consuming. An efficient model with high positioning accuracy is therefore not yet available. We propose a stacking ensemble broad learning localization system using channel state information as a fingerprint, which is termed EnsemLoca. A bootstrapping method is used to build the training set, which enables the EnsemLoca system to build the base learner in parallel by using bagging. The broad learning system (BLS), which is a novel neural network model, as a base learner, not only has the advantage of time complexity but also offers a sparse representation in which the features are filtered. A unique base learner is constructed by randomly selecting the samples and features, and they are combined by stack generalization. The experimental results show that the EnsemLoca system achieves higher accuracy than several machine-learning algorithms in both line-of-sight (LOS) and non-LOS environments, and is even stronger than deep neural networks characterized by accuracy. At the same time, it has the same theoretical complexity as BLS, which greatly reduces the offline training time.

In Mobile crowdsensing (MCS), the platform needs an adequate user group to accomplish tasks. Its recruiting worker strategy is essential for sensor data quality. Social-network-assisted worker recruitment effectively expands task coverage. However existing studies enclose two impractical assumptions: influence between users is determined by the number of friends and the recruiting reward is fixed. To solve this problem, a novel influence and cost trade-off (ICT) algorithm is proposed to apply the worker recruitment strategy in the real world. ICT uses linear equations to estimate influence and cost iteratively under the impact of the seed set. Using the influence model based on social interaction, the algorithm selects a near-optimal set of seeds by the revenue-cost-ratio. Empirical studies on three realworld datasets verify that ICT achieves higher performance than baseline methods under various settings.

Good user experiences in Mobile Augmented Reality (MAR) applications require timely processing and rendering of virtual objects on user devices. Today's wearable AR devices are limited in computation, storage, and battery lifetime. Edge computing, where edge devices are employed to offload part or all computation tasks, allows an acceleration of computation without incurring excessive network latency. In this paper, we use acyclic data flow graphs to model the computation and data flow in MAR applications and aim to minimize the makespan of processing input frames. Due to task dependencies and variable resource availability, makespan minimization is proven to be NP-hard in general. We design DPA, a polynomial-time heuristic algorithm for this problem. For special data flow graphs including chain or star, the algorithm can provide optimal solutions or solutions with a constant approximation ratio. The effectiveness of DPA has been evaluated using extensive simulations with realistic workloads and resource availability measured from a prototype implementation.

Cloud storage services such as Dropbox have been widely used for file collaboration among multiple users. However, this desirable functionality is yet restricted to the “walled-garden” of each service. At present, the only feasible approach to cross-cloud file collaboration seems to be using web APIs, whose performance is known to be highly unstable and unpredictable. Now that using inefficient web APIs is inevitable, in this paper we attempt to achieve sound user-perceived performance for cross-cloud file collaboration. This attempt is enabled by two key observations from real-world measurements. First, for each cloud, we are always able to deploy one or several nearby (client) proxies which can efficiently access the web APIs. Second, during file collaboration, significant similarity exists among different versions of a file. This can be exploited to substantially reduce inter-proxy traffic and thus shorten the data sync time. Guided by the observations, we design and implement an open-source prototype system called CoCloud. Currently, it supports file collaboration among four popular cloud storage services in the US and China. Its performance is well acceptable to users under representative workloads, even approaching or exceeding that of intra-cloud collaboration in many cases.

The exhaustion of IPv4 addresses has led to the transition to IPv6 becoming a major task for the Internet. Providing IPv6 users with accessibility to IPv4 services has been one of the most challenging tasks during the IPv4/IPv6 transition. However, some protocols contain IP addresses in the application layer, resulting in incompatibilities in traversing the network-layer translators. In this paper, we propose a Generic Application Layer Translator (GALT), which is designed to bridge the gap between IPv6 clients and IPv4 cross-layer services for the IPv4/IPv6 translation scenario. To ensure correctness and efficiency, we propose a protocol description language that enables GALT to be aware of protocol semantics. We develop the GALT system and apply it to widely used protocols with cross-layer issues, including HTTP and SIP. Our evaluation shows that GALT correctly solves the cross-layer problem in IPv4/IPv6 translation, with efficient performance for practical services.

Cloud storage services such as Dropbox have been widely used for file collaboration among multiple users. However, this desirable functionality is yet restricted to the “walled-garden” of each service. At present, the only effective approach to cross-cloud file collaboration seems to be using web APIs, whose performance is known to be highly unstable and unpredictable. Now that using inefficient web APIs is inevitable, in this paper we attempt to achieve sound user-perceived performance for cross-cloud file collaboration. This attempt is enabled by two key observations from real-world measurements. First, for each cloud, we are always able to deploy one or several nearby (client) proxies which can efficiently access the web APIs. Second, during file collaboration, significant similarity exists among different versions of a file. This can be exploited to substantially reduce inter-proxy traffic and thus shorten the data sync time. Guided by the observations, we design and implement an open-source prototype system called CoCloud. Currently, it supports file collaboration among four popular cloud storage services in the US and China. Its performance is well acceptable to users under representative workloads, even approaching or exceeding intra-cloud performance in many cases.

In this paper, we investigate the scheduling problem with multi-resource allocation in cloud computing environments. In contrast to existing work that focuses on flow-level scheduling, which treats flows in isolation, we consider dependency among subtasks of applications that imposes a partial order relationship in execution. We formulate the problem of Multi-Resource Partial-Ordered Task Scheduling (MR-POTS) to minimize the makespan. In the first stage, the proposed Dominant Resource Priority (DRP) algorithm decides the collection of subtasks for resource allocation by taking into account the partial order relationship and characteristics of subtasks. In the second stage, the proposed Maximum Utilization Allocation (MUA) algorithm partitions multiple resources among selected subtasks with the objective to maximize the overall utilization. Both theoretical analysis and experimental evaluation demonstrate the proposed algorithms can approximately achieve the minimal makespan with high resource utilization. Specifically, a reduction of 50% in makespan can be achieved compared with existing scheduling schemes.