融合重检测机制的卷积回归网络目标跟踪算法

针对基于人工特征的背景感知相关滤波（CACF）算法在形变、运动模糊、低分辨率情形跟踪效果较差以及跟踪器遇到严重遮挡等情形容易陷入局部最优而导致跟踪失败的问题，提出一种融合重检测机制的卷积回归网络（CRN）目标跟踪算法。在训练阶段，将相关滤波作为CRN层融入进深度神经网络，使网络成为一个整体进行端到端训练；在跟踪阶段，通过残差连接融合不同网络层及其响应值，同时引入重检测机制使算法从潜在的跟踪失败中恢复，当响应值低于给定阈值时激活检测器。在数据集OTB-2013上的实验表明，所提算法在50个视频序列上精确度达到88.1%，相比原始CACF算法提高9.7个百分点，在具有形变、运动模糊等属性的视频序列上相比原始算法表现更优秀。     

高斯混合模型结合加权似然的目标跟踪算法

鉴于高斯混合模型对背景变化快时无法精确检测出目标和目标跟踪的适应性差等瑕疵，提出了基于加权似然跟踪器来改进高斯混合模型实现运动目标跟踪算法。主要引入了自适应高斯混合模型来实时检测运动目标，然后空间加权似然来进行视频中的目标定位，引入加权似然期望值来改进高斯混合模型处理视频中的多尺度、多角度变化的目标跟踪不精准问题。通过VOT 2014 dataset对比实验结果表明提出的基于加权似然跟踪（Weighted Likelihood Tracking，WLT）和改进高斯混合模型（Improved Gaussian Mixture Model，IGMM）的目标跟踪算法较传统高斯混合模型跟踪算法在跟踪的精度有较大提高。在应对多尺度、多角度变化的目标跟踪表现出了较大的优势。     

Adaptive finite‐time formation tracking control for multiple nonholonomic UAV system with uncertainties and quantized input

An adaptive finite‐time formation tracking control approach is proposed for multiple unmanned aerial vehicle (UAV) system with quantized input signals in this paper. The UAVs are described by nonholonomic kinematic model and autopilot model with uncertainties. An enhanced hysteretic quantizer is introduced to avoid chattering, and some restrictions are released by using a new quantization decomposition method. Based on backstepping technique and finite‐time Lyapunov stability theory, the adaptive finite‐time controller is designed for the trajectory tracking of the multi‐UAV formation. The nonholonomic constraints are solved by a transverse function. A transformation is introduced to the control input signals to eliminate the quantization effect. Stability analysis proves that the tracking errors can converge to a small neighborhood of the origin within finite time and all the closed‐loop signals are semiglobally finite‐time bounded. The effectiveness of the proposed control approach is validated by simulation and experiment.     

车辆高速避让路径跟踪控制策略

建立车辆动力学模型,研究车辆在高速避让工况下的路径跟踪控制策略。基于自抗扰控制设计车辆横摆角速度以及道路曲率2种不同的路径跟踪控制器;设计基于车辆偏航位移和偏航角度加权的线性二次最优控制器;对3种控制器控制效果进行仿真验证,分析路径跟踪侧向位移误差、侧向加速度、前轮转向角以及偏航角度变化。越野车实车测试结果表明,路径跟踪控制策略正确。     

A multi-tone central divided difference frequency tracker with adaptive process noise covariance tuning

The problem of real-time frequency estimation of nonstationary multi-harmonic signals is important in many applications. In this paper, we propose a novel multi-frequency tracker based on a state-space representation of the signal with Cartesian filters and the second-order central divided difference filter (CDDF), which improves the performance of the extended Kalman filter (EKF) by using Stirling's interpolation method to approximate the mean and covariance of the state vector. A crucial element of the method is the adaptive scaling of the process noise covariance matrix appearing in the filter equations, as a function of the innovation sequence, which tunes the accuracy-reactivity trade-off of the filter. The proposed solution is evaluated against two approaches from the literature, namely the factorized adaptive notch filter (FANF) and the extended Kalman filter frequency tracker (EKFFT). Several experiments emphasize the estimation accuracy of the proposed method as well as the improved robustness with respect to initial errors and input signal complexity. The presented method appears to be particularly efficient with rapidly varying frequencies, thanks to the update mechanism that adjusts the filter parameters based on the amplitude of the estimation error.     

Basic human–robot interaction system running on an embedded platform

Robotics will be a dominant area in society throughout future generations. Its presence is currently increasing in most daily life settings, with devices and mechanisms that facilitate the accomplishment of diverse tasks, as well as in work scenarios, where machines perform more and more jobs. This increase in the presence of autonomous robotic systems in society is due to their great efficiency and security compared to human capacity, which is thanks mainly to the enormous precision of their sensor and actuator systems. Among these, vision sensors are of the utmost importance. Humans and many animals naturally enjoy powerful perception systems, but, in robotics, this constitutes a constant line of research. In addition to having a high capacity for reasoning and decision-making, these robots incorporate important advances in their perceptual systems, allowing them to interact effectively in the working environments of this new industrial revolution. Drawing on the most basic interaction between humans, looking at the face, an innovative system is presented in this paper, which was developed for an autonomous and DIY robot. This system is composed of three modules. First, the face detection component, which detects human faces in the current image. Second, the scene representation algorithm, which offers a wider field of view than that of the single camera used, mounted on a servo-pan unit. Third, the active memory component, which was designed and implemented according to two competing dynamics: life and salience. The algorithm intelligently moves the servo-pan unit with the aim of finding new faces, follow existing ones and forgetting those that no longer appear on the scene. The system was developed and validated using a low-cost platform based on a Raspberry Pi3 board.     

Decentralized optimal tracking control for large-scale nonlinear systems with tracking error constraints

In this article, a decentralized optimal tracking control strategy is proposed for a class of nonlinear systems with tracking error constraints by utilizing adaptive dynamic programming (ADP). It should be noted that ADP technology cannot be directly used to solve decentralized optimal tracking problem of large-scale interconnected nonlinear system with nonzero equilibrium points, since that an infinite domain performance index function may result in an unsolvable solution. In addition, by introducing a smooth function, the constrained tracking error is transformed into an unconstrained one. Then, the error dynamics and a new infinite domain performance index function are designed, such that ADP technology can be used. Following the designed performance index function, the tracking error can be ensured within a small neighborhood of zero. Finally, the feasibility and the effectiveness of the proposed decentralized optimal control scheme are verified through two simulation examples.     

Multi-Level Fusion Net for hand pose estimation in hand-object interaction

This work is about solving a challenging problem of estimating the full 3D hand pose when a hand interacts with an unknown object. Compared to isolated single hand pose estimation, occlusion and interference induced by the manipulated object and the clutter background bring more difficulties for this task. Our proposed Multi-Level Fusion Net focuses on extracting more effective features to overcome these disadvantages by multi-level fusion design with a new end-to-end Convolutional Neural Network (CNN) framework. It takes cropped RGBD data from a single RGBD camera at free viewpoint as input without requiring additional hand–object pre-segmentation and object or hand pre-modeling. Through extensive evaluations on public hand–object interaction dataset, we demonstrate the state-of-the-art performance of our method.     

瓷砖表面非常规尺寸瑕疵的检测

与普通目标检测任务不同, 瓷砖表面瑕疵检测的困难之处在于检测小尺寸和大长宽比等非常规尺寸的目标. 为了解决这两个问题, 本文提出了一种基于改进 Cascade R-CNN 的新型瓷砖表面瑕疵检测算法. 为了提高模型对小瑕疵的检测能力, 本文模型利用侧向连接结构进行上下层语义信息的融合, 使用可切换空洞率的空洞卷积来增加模型的感受野; 为了提高模型对于大长宽比瑕疵的检测能力, 本文模型在标准卷积上引入偏移域以更好提取目标特征信息. 此外, 本文模型调整 Cascade R-CNN 框架中预选锚框的大小和长宽比例. 实验结果表明, 在从瓷砖工厂收集的数据集上, 本文所提出算法的平均精度均值 (mean average precision, mAP) 达到了 73.5%, 比改进前的 Cascade R-CNN 模型提高了9.7%. 本文实验代码可从以下链接获取: https://github.com/mashibin/Ceramic-tile-defect-detection.