摘要：
自动驾驶车辆在行驶过程中,需要对行人和车辆同时完成目标检测、实例分割和目标跟踪三个任务.提出一种基于深度学习的环境感知模型同时对三个任务进行多任务学习.首先,通过卷积神经网络对连续帧图像提取时空特征;然后,通过注意力机制对时空特征进行去耦再融合,充分利用任务间的相关性,实现不同任务对时空特征的差异化选择;最后,为平衡不同任务间的学习速率,使用动态加权平均的方式对模型进行训练.在KITTI数据集上的实验结果表明,所提模型在目标检测方面,比CenterTrack模型F1得分提高了0.6个百分点;在目标跟踪方面,比TraDeS(Track to Detect and Segment)模型多目标跟踪精度(MOTA)提高了0.7个百分点;在实例分割方面,比SOLOv2（Segmenting Objects by LOcations version 2）模型AP50 和AP75 分别提高了7. 4和3. 9个百分点.
作者机构：
[Liu, Shouyin; He, MJ; He, Mingjian] Cent China Normal Univ, Coll Phys Sci & Technol, Wuhan 430079, Hubei, Peoples R China.
通讯机构：
[He, MJ ] C;Cent China Normal Univ, Coll Phys Sci & Technol, Wuhan 430079, Hubei, Peoples R China.
摘要：
Continuous variable quantum teleportation provides a path to the long-distance transmission of quantum states. Photon-varying non-Gaussian operations have been shown to improve the fidelity of quantum teleportation when integrated into the protocol. However, for a given type of non-Gaussian operation, the achievable fidelity varies with the parameters associated with the operation. Previous work only focused on particular settings of the parameters, over which an optimization was missing. The potential of such operations is not fully uncovered. Given a fixed non-Gaussian operation, the achievable fidelity also varies with input states. An operation that increases the fidelity for teleporting one class of states might do the contrary for other classes of states. A performance metric, upon which an operation is optimized, suitable for different input states, is also missing. In this work, we build a framework for photon-varying non-Gaussian operations for multimode states, upon which we propose a performance metric suitable for arbitrary teleportation input states. We then apply the new metric to evaluate different types of non-Gaussian operations. Starting from simple multiphoton photon subtraction and photon addition, we find that increasing the number of ancillary photons involved in the operation does not guarantee performance improvement. We then investigate a generalization of the operations mentioned above, finding that operations that approximate a particular form provide the best improvement. The results provided here will be valuable for real-world implementations of quantum teleportation networks and applications that harness the non-Gaussianity of quantum states.
摘要：
Wireless propagation models play a significant role in the deployments of base stations that are used to the reference signal receiving power (RSRP) of signal receivers in a cell. However, the existing models predict the RSRP of one receiver point in a cell at a time, which cannot be generalized to other cells. Motivated by this, a cell-level RSRP estimation method is proposed to directly predict the whole-cell RSRP by converting the RSRP estimation into an image-to-image translation. First, an environment map of each cell and measured RSRP for each cell is transformed into an image. Second, a cell-level image-to-image wireless propagation model based on conditional generative adversarial networks is proposed, which can directly predict the whole-cell RSRP at a time. In particular, a residual estimation method is proposed for the measurement RSRP data in the real world. The proposed method employs an empirical model to reveal the wireless propagation law as a priori knowledge and guide the training steps of the deep learning model. Finally, the experimental results verify the accuracy and generalization performance of the proposed image-to-image wireless propagation model.
作者机构：
[Wang, Ji; Liu, Shouyin; Xiao, Jian] Cent China Normal Univ, Dept Elect & Informat Engn, Wuhan, Peoples R China.;[Liu, Yuanwei] Queen Mary Univ London, Sch Elect Engn & Comp Sci, London, England.;[Xie, Wenwu] Hunan Inst Sci & Technol, Sch Informat Sci & Engn, Yueyang, Peoples R China.;[Wang, Jun] CICT Mobile Commun Technol Co Ltd, Wuhan, Peoples R China.
会议名称：
GLOBECOM 2023 - 2023 IEEE Global Communications Conference
会议时间：
04 December 2023
会议地点：
Kuala Lumpur, Malaysia
会议主办单位：
[Xiao, Jian;Wang, Ji;Liu, Shouyin] Cent China Normal Univ, Dept Elect & Informat Engn, Wuhan, Peoples R China.^[Liu, Yuanwei] Queen Mary Univ London, Sch Elect Engn & Comp Sci, London, England.^[Xie, Wenwu] Hunan Inst Sci & Technol, Sch Informat Sci & Engn, Yueyang, Peoples R China.^[Wang, Jun] CICT Mobile Commun Technol Co Ltd, Wuhan, Peoples R China.
会议论文集名称：
GLOBECOM 2023 - 2023 IEEE Global Communications Conference
摘要：
A joint cascaded channel estimation framework is proposed for simultaneously transmitting and reflecting reconfigurable intelligent surfaces (STAR-RIS) systems with hardware imperfection, in which practical the hybrid-field electromagnetic wave radiation with spatial non-stationarity is investigated. By exploiting the cascaded channel correlations in user domain and STAR-RIS element domain, we propose a multi-task network (MTN) with multi-expert branches to simultaneously reconstruct the high-dimensional transmitting and reflecting channels from the observed mixture channel with noise. In the proposed MTN architecture, a learnable shrinkage module is exploited to constrict the communication noise, and self-attention mechanism-based Transformer layers are utilized to extract the non-local feature of the non-stationary cascaded channel. Numerical results show that the proposed MTN achieves superior channel estimation accuracy with less training overhead compared with existing state-of-the-art benchmarks, in terms of required pilots, computations, and network parameters.
作者机构：
[Ji, Wang; Liu, Shouyin; Yi, Zheng; Rong, Huang] Cent China Normal Univ, Coll Phys Sci & Technol, Wuhan 430079, Hubei, Peoples R China.;[Liu, Zhiwen] Cent China Normal Univ, Wollongong Joint Inst, Wuhan 430079, Hubei, Peoples R China.;[Xie, Wenwu] Hunan Inst Sci & Technol, Yueyang 414006, Peoples R China.
通讯机构：
[Shouyin, L.] C;College of Physical Science and Technology, China
关键词：
convolution neural networks;Feature extraction;reference signal received power
摘要：
In this letter, an environmental features (EFs) extraction model is proposed for estimating reference signal received power (RSRP) accurately. Firstly, 18-D measured data is transformed into 15-D physical features (PFs). Then 15-D PFs is reduced to 14-D by performing correlation analysis. Secondly, EFs are extracted from the environmental maps (EMs) by applying Convolution Neural Networks (CNNs). Finally, several Machine Learning Regressors (MLRs) are trained to predict RSRP combining PFs and EFs as inputs. The results, in test dataset, show that prediction performance of MLRs is improved through 14-D PFs, and is further improved in nonlinear MLRs combining PFs and EFs.
作者机构：
[Fu, Baochi; Song, Huichao] Peking Univ, Dept Phys, Beijing 100871, Peoples R China.;[Fu, Baochi; Song, Huichao] Peking Univ, State Key Lab Nucl Phys & Technol, Beijing 100871, Peoples R China.;[Fu, Baochi; Song, Huichao] Collaborat Innovat Ctr Quantum Matter, Beijing 100871, Peoples R China.;[Liu, Shuai Y. F.; Yin, Yi] Chinese Acad Sci, Quark Matter Res Ctr, Inst Modern Phys, Lanzhou 730000, Peoples R China.;[Pang, Longgang] Cent China Normal Univ, Key Lab Quark & Lepton Phys MOE, Wuhan 430079, Peoples R China.
通讯机构：
[Shuai Y. F. Liu] Q;Quark Matter Research Center, Institute of Modern Physics, Chinese Academy of Sciences, Lanzhou 730000, China
摘要：
We study the spin polarization generated by the hydrodynamic gradients. In addition to the widely studied thermal vorticity effects, we identify an undiscovered contribution from the fluid shear. This shear-induced polarization (SIP) can be viewed as the fluid analog of strain-induced polarization observed in elastic and nematic materials. We obtain the explicit expression for SIP using the quantum kinetic equation and linear response theory. Based on a realistic hydrodynamic model, we compute the differential spin polarization along both the beam direction (z) over cap and the out-plane direction (y) over cap in noncentral heavy-ion collisions at root s(NN) = 200 GeV, including both SIP and thermal vorticity effects. We find that SIP contribution always shows the same azimuthal angle dependence as experimental data and competes with thermal vorticity effects. In the scenario that Lambda inherits and memorizes the spin polarization of a strange quark, SIP wins the competition, and the resulting azimuthal angle dependent spin polarization P (y) and P- z agree qualitatively with the experimental data.
