Object localization has a vital role in any object detector and tracker, and therefore, has been the focus of attention by many researchers. In this article, a special training approach is proposed for a light convolutional neural network (CNN) to determine the region of interest (RoI) in an image while effectively reducing the number of probable anchor boxes. Almost all CNN based detectors utilize a fixed input size image, which may yield poor performance when dealing with various object sizes. In this paper, a different CNN structure is proposed taking three different input sizes, to enhance the performance. To demonstrate the effectiveness of the proposed method, two common data set are used for training while tracking by localization application is considered to demonstrate its final performance. The promising results indicate the applicability of the presented structure and the training method in practice.

Object tracking becomes critical especially when similar objects are present in the same area. Recent state-of-the-art (SOTA) approaches are proposed based on taking a matching network with a heavy structure to distinguish the target from other objects in the area which indeed drastically downgrades the performance of the tracker in terms of speed. Besides, several candidates are considered and processed to localize the intended object in a region of interest for each frame which is time-consuming. In this article, a special architecture is proposed based on which in contrast to the previous approaches, it is possible to identify the object location in a single shot while taking its template into account to distinguish it from the similar objects in the same area. In brief, first of all, a window containing the object with twice the target size is considered.

Both recognition and 3D tracking of frontal dynamic objects are crucial problems in an autonomous vehicle, while depth estimation as an essential issue becomes a challenging problem using a monocular camera. Since both camera and objects are moving, the issue can be formed as a structure from motion (SFM) problem. In this paper, to elicit features from an image, the YOLOv3 approach is utilized beside an OpenCV tracker. Subsequently, to obtain the lateral and longitudinal distances, a nonlinear SFM model is considered alongside a state-dependent Riccati equation (SDRE) filter and a newly developed observation model.

As a critical part of any security system, identity recognition has become paramount among researchers. In this regard, several methods are presented while considering various sensors and data. In particular, gait data yields rich information about a person, including some exclusive moving patterns which can be utilized to distinguish between different individuals. On the other hand, convolutional neural networks are proved to be applicable for structured data, especially images. In this article, 12 markers are considered in gathering the gait data, each representing a lower-body joint location. Then, utilizing the gait data in a 2D tensor form, three different convolutional neural networks are trained to recognize the identities. Taking light architectures into account, this approach is implementable in realtime application. The obtained result shows the promising capability of the proposed method being used in identity …

Human Activity Recognition (HAR) is a challenging problem that needs advanced solutions than using handcrafted features to achieve a desirable performance. Deep learning has been proposed as a solution to obtain more accurate HAR systems being robust against noise. In this paper, we introduce ARC-Net and propose the utilization of capsules to fuse the information from multiple inertial measurement units (IMUs) to predict the activity performed by the subject. We hypothesize that this network will be able to tune out the unnecessary information and will be able to make more accurate decisions through the iterative mechanism embedded in capsule networks. We provide heatmaps of the priors, learned by the network, to visualize the utilization of each of the data sources by the trained network. By using the proposed network, we were able to increase the accuracy of the state-of-the-art approaches by 2%. Furthermore, we investigate the directionality of the confusion matrices of our results and discuss the specificity of the activities based on the provided data.

Lithium-ion (Li-ion) batteries need very precise monitor of the state of charge (SOC) to ensure a long cycle life.
Hence, a knowledge of the SOC is important for Li-ion batteries. Although SOC cannot be measured directly, it can be estimated from direct measurement variables based on a model of the battery. Single-Particle-Model (SPM), a reduced-order nonlinear electrochemical model, is commonly used for this purpose. State-dependent-
Riccati-equation (SDRE) filter is chosen as the estimator due to its high-flexibility in handling the model’s nonlinearity. However, performance of this filter is limited in presence of uncertainties. To tackle this problem,
in this paper, a switching concept is induced into SDRE filter, in the form of switched estimation error covariance
matrix with a certain frequency. Thus, by changing the Riccati equation dynamic in SDRE filter and proper
adjustment of estimation error covariance matrix eigenvalues, performance and robustness of the common SDRE
filter is significantly improved for Li-ion SOC estimation. To analyze the fidelity of such a filter in further
applications, stability analysis is carried out on a class of nonlinear systems, and ultimate bound of estimation error is analytically obtained, and the influence of switching is investigated. Simulation results reveal effec-
tiveness of the proposed filter compared to common SDRE filter, extended Kalman filter and variable structure approaches. Furthermore, experimental results verify the effectiveness of the proposed method in practice.

Human Activity Recognition (HAR) is a challenging problem that needs advanced solutions than using handcrafted features to achieve a desirable performance. Deep learning has been proposed as a solution to obtain more accurate HAR systems being robust against noise. In this paper, we introduce ARC-Net and propose the utilization of capsules to fuse the information from multiple inertial measurement units (IMUs) to predict the activity performed by the subject. We hypothesize that this network will be able to tune out the unnecessary information and will be able to make more accurate decisions through the iterative mechanism embedded in capsule networks. We provide heatmaps of the priors, learned by the network, to visualize the utilization of each of the data sources by the trained network. By using the proposed network, we were able to increase the accuracy of the state-of-the-art approaches by 2%. Furthermore, we investigate the directionality of the confusion matrices of our results and discuss the specificity of the activities based on the provided data.

Object tracking through image sequences is one of the important components of many vision systems, and it has numerous applications in driver assistance systems such as pedestrian collision avoidance or collision mitigating systems. Blurred images produced by a rolling shutter camera or occlusions may easily disturb the object tracking system. In this article, a method based on convolutional and recurrent neural networks is presented to further enhance the performance and robustness of such trackers. It is proposed to use a convolutional neural network to detect an intended object and feed the tracker with found image. Moreover, by using this structure the tracker is updated every ' n ' frames. A recurrent neural network is designed to learn the object behavior for estimating and predicting its position in blurred frames or when it is occluded behind an obstacle. Real-time implementation of the proposed approach verifies its applicability for improvement of the trackers performance.

Multi-goal motion planning under motion and sensor uncertainty is the problem of finding a reliable policy for visiting a set of goal points. In this paper, the problem is formulated as a formidable traveling salesman problem in the belief space. To solve this intractable problem, we propose an algorithm to construct a TSP-FIRM graph which is based on the feedback-based information roadmap (FIRM) algorithm. Also, two algorithms are proposed for the online planning of the obtained policy in the offline mode and overcoming changes in the map of the environment. Finally, we apply the algorithms on a physical nonholonomic mobile robot in the presence of challenging situations like the discrepancy between the real and computation model, map updating and kidnapping.

Representation of the surrounding environment is a vital task for a mobile robot. Many applications for mobile robots in urban environments may be considered such as self-driving cars, delivery drones or assistive robots. In contrast to the conventional methods, in this paper a Non Uniform Rational B-Spline (NURBS) based technique is represented for 3D mapping of the surrounding environment. While in the state of the art techniques, the robot's environment is expressed in a discrete space, the proposed method is mainly developed for representation of environment in a continuous space. Exploiting the information theory, the generated representation has much lower complexity and more compression capability in relation to some state of the art techniques. In addition to representation in a lower dimensional space, the NURBS based representation is invariant against 3D geometric transformations. Furthermore, the NURBS based representation can be employed for obstacle avoidance and navigation. The applicability of the proposed algorithm is investigated in some urban environments through some publicly available data sets. It has been shown by some experiments that the proposed method has better visual representation and much better data compression compared to some state-of-the-art methods.

In this study, the problem of navigation in dynamic and unknown environment is investigated and a navigation method based on force field approach is suggested. It is assumed that the robot performs navigation in unknown environment and builds the map through SLAM procedure. Since the moving objects' location and properties are unknown, they are identified and tracked by Kalman filter. Kalman observer provides important information about next paths of moving objects which are employed in finding collision point and time in future. In the time of collision detection, a modifying force is added to repulsive and attractive forces corresponding to the static environment and leads the robot to avoid collision. Moreover, a safe turning angle is defined to assure safe navigation of the robot. The performance of proposed method, named Escaping Algorithm, is verified through different simulation and experimental tests. Besides, comparison between Escaping Algorithm and Probabilistic Velocity Obstacle, based on computational complexity and required steps for finishing the mission is provided in this paper. The results show Escaping Algorithm outperforms PVO in term of dynamic obstacle avoidance and complexity as a practical method for autonomous navigation.

In the problem of simultaneously localization and mapping (SLAM) for a mobile robot, it is required to detect previously visited locations so the estimation error shall be reduced. Sensor observations are compared by a similarity metric to detect loops. In long term navigation or exploration, the number of observations increases and so the complexity of the loop closure detection. Several techniques are proposed in order to reduce the complexity of loop closure detection. Few algorithms have considered the loop closure detection from a subset of sensor observations. In this paper, the compressed sensing approach is exploited to detect loops from few sensor measurements. In the basic compressed sensing it is assumed that a signal has a sparse representation is a basis which means that only a few elements of the signal are non-zero. Based on the compressed sensing approach a sparse signal can be recovered from few linear noisy projections by l1 minimization. The difference matrix which is widely used for loop detection has a sparse structure, where similar observations are shown by zero distance and different locations are indicated by ones. Based on the multiple measurement vector technique which is an extension of the basic compressed sensing, the loop closure detection is performed by comparison of few sensor observations. The applicability of the proposed algorithm is investigated in some outdoor environments through some publicly available data sets. It has been shown by some experiments that the proposed method can detect loops effectively.

In this paper the problem of 3D scene and object classification from depth data is addressed. In contrast to high-dimensional feature-based representation, the depth data is described in a low dimensional space. In order to remedy the curse of dimensionality problem, the depth data is described by a sparse model over a learned dictionary. Exploiting the algorithmic information theory, a new definition for the Kolmogorov complexity is presented based on the Earth Mover’s Distance (EMD). Finally the classification of 3D scenes and objects is accomplished by means of a normalized complexity distance, where its applicability in practice is proved by some experiments on publicly available datasets. Also, the experimental results are compared to some state-of-the-art 3D object classification methods. Furthermore, it has been shown that the proposed method outperforms FAB-Map 2.0 in detecting loop closures, in the sense of the precision and recall.

This paper presents an online robust RGB-D SLAM algorithm which uses an improved switchable constraints robust pose graph slam alongside with radial variance based hash function as the loop detector. The switchable constraints robust back-end is improved by initialization of its weights according to information matrix of the loops and is validated using real world datasets. The radial variance based hash function is combined with an online image to map comparison to improve accuracy of loop detection. The whole algorithm is implemented on K. N. Toosi University mobile robot with a Microsoft Kinect camera as the RGB-D sensor and the whole algorithm is validated using this robot, while the map of the environment is generated in an online fashion.

An improved square root unscented fast simultaneous localization and mapping (FastSLAM) is proposed in this paper. The proposed method propagates and updates the square root of the state covariance directly in Cholesky decomposition form. Since the choice of the proposal distribution and that of the resampling method are the most critical issues to ensure the performance of the algorithm, its optimization is considered by improving the sampling and resampling steps. For this purpose, particle swarm optimization (PSO) is used to optimize the proposal distribution. PSO causes the particle set to tend to the high probability region of the posterior before the weights are updated; thereby, the impoverishment of particles can be overcome. Moreover, a new resampling algorithm is presented to improve the resampling step. The new resampling algorithm can conquer the defects of the resampling algorithm and solve the degeneracy and sample impoverishment problem simultaneously. Compared to unscented FastSLAM (UFastSLAM), the proposed algorithm can maintain the diversity of particles and consequently avoid inconsistency for longer time periods, and furthermore, it can improve the estimation accuracy compared to UFastSLAM. These advantages are verified by simulations and experimental tests for benchmark environments.

FastSLAM is a framework for simultaneous localization and mapping (SLAM). However, FastSLAM algorithm has two serious drawbacks, namely the linear approximation of nonlinear functions and the derivation of the Jacobian matrices. For solving these problems, UFastSLAM has been recently proposed. However, UFastSLAM is inconsistent over time due to the loss of particle diversity that is caused mainly by the particle depletion in the resampling step and incorrect a priori knowledge of process and measurement noises. To improve consistency, intelligent UFastSLAM with Markov chain Monte Carlo (MCMC) move step is proposed. In the proposed method, the adaptive neuro-fuzzy inference system supervises the performance of UFastSLAM. Furthermore, the particle impoverishment caused by resampling is restrained after the resample step with MCMC move step. Simulations and experiments are presented to evaluate the performance of algorithm in comparison with UFastSLAM. The results show the effectiveness of the proposed method.

This paper addresses a new method for navigation in dynamic environment. The proposed method is based on force field method and it is supposed that the robot performs SLAM and autonomous navigation in dynamic environment without any predefined information about dynamic obstacles. The movement of dynamic obstacles is predicted by Kalman filter and is used for collision detection purpose. In the time of collision detection, a modifying force is added to repulsive and attractive forces corresponding to the static environment and leads robot to avoid collision. Moreover, a safe turning angle is defined to assure safe navigation of the robot. The performance of proposed method, named Escaping Algorithm, is verified through different simulation and experimental tests. The results show the proper performance of Escaping Algorithm in term of dynamic obstacle avoidance as a practical method for autonomous navigation.

FastSLAM is a framework for simultaneous localisation and mapping (SLAM) using a Rao-Blackwellised particle filter. In FastSLAM, particle filter is used for the robot pose (position and orientation) estimation, and parametric filter (i.e. EKF and UKF) is used for the feature location's estimation. However, in the long term, FastSLAM is an inconsistent algorithm. In this paper, a new approach to SLAM based on hybrid auxiliary marginalised particle filter and differential evolution (DE) is proposed. In the proposed algorithm, the robot pose is estimated based on auxiliary marginal particle filter that operates directly on the marginal distribution, and hence avoids performing importance sampling on a space of growing dimension. In addition, static map is considered as a set of parameters that are learned using DE. Compared to other algorithms, the proposed algorithm can improve consistency for longer time periods and also, improve the estimation accuracy. Simulations and experimental results indicate that the proposed algorithm is effective.

In this paper a new controller for nonholonomic system is introduced. This feedback error learning controller benefits from both nonlinear and adaptive controller properties. The nonlinear controller is used to stabilize the nonholonomic behavior of the systems. This controller is a sliding mode controller which is designed based on backstepping method. The adaptive controller tries to face with uncertainty and unknown dynamic of the mobile robot. This part uses neural network controller for adaptation. The experimental results show the effectiveness of proposed controller and suitable and robust tracking performance of a mobile robot, which is significantly better than traditional controllers.

This paper proposes a method for mobile robot exploration based on the idea of frontier exploration which suggests navigating the robot toward the boundaries between free and unknown areas in the map. A global occupancy grid map of the environment is constantly updated, based on which a global frontier map is calculated. Then, a histogram based approach is adopted to cluster frontier cells and score these clusters based on their distance from the robot as well as the number of frontier cells they contain. In each stage of the algorithm, a sub-goal is set for the robot to navigate. A combination of distance transform and A* search algorithms is utilized to generate a plausible path toward the sub-goal through the free space. This way keeping a reliable distance from obstacles is guaranteed while searching for the shortest path toward the sub-goal. When such a path is generated, a B-spline interpolated and smoothed trajectory is produced as the control reference for the mobile robot to follow. The whole process is iterated until no unexplored area remains in the map. The efficiency of the method is shown through simulated and real experiments.

FastSLAM is a framework using a Rao-Blackwellized particle filter. However, the performance of FastSLAM depends on correct a priori knowledge of the process and measurement noise covariance matrices (Q t and R t ) that are in most applications unknown. On the other hand, an incorrect a priori knowledge of Q t and R t may seriously degrade the performance of FastSLAM. To solve these problems, this paper presents H ? FastSLAM. In this approach, H-Infinity particle filter is used for the mobile robot position estimation and H-Infinity filter is used for the feature location's estimation. The H-Infinity FastSLAM can work in an unknown statistical noise behavior and thus it is more robust. Experimental results demonstrate the effectiveness of the proposed algorithm.

We introduce a very fast and robust localization and 2D environment representation algorithm in this paper. This innovative method matches lines extracted from the LASER range finder distance data with the lines that construct the map, in order to calculate the local translation and rotation. This matching is done with a simple least square with no iterations. The algorithm is suitable for any indoor environment with mostly polygonal structure and has proven high speed and robustness in the experimental tests on our innovatively designed tracked mobile rescue robot ldquoSilverrdquo. One experimental test is presented in the last section of this paper where the outputs are presented. These outputs are: 1-drift free raster map made of points and 2- A gallery of lines providing a linear ground truth.

In this paper, a novel hybrid fuzzy proportional-integral-derivative (PID) controller based on learning automata for optimal tracking of robot systems including motor dynamics is presented. Learning automata is used at the supervisory level for adjustment of the parameters of hybrid Fuzzy-PID controller during the system operation. The proposed method has better convergence rate in comparison with standard back-propagation algorithms, less computational requirements than adaptive network based fuzzy inference systems (ANFIS) or neural based controllers and having the ability of working in uncertain environments without any previous knowledge of environments' parameters. The proposed controller has been successfully applied in simulation to control a 6-DOF Puma 560 manipulator using robotic toolbox, and has satisfactory results. In this simulation also, external disturbance and noise are addressed. The result of simulation has also shown that the rate of convergence and robustness of the designed controller guarantees practical stability.

Presented method in this paper aims to develop an accurate motion model and SLAM algorithm, which is only based on the Laser Range Finder (LRF) data. Proposed method tries to overcome some practical problems in traditional motion models and SLAM approaches, such as robot slippage, and inaccuracy in parameters related to robot's hardware. Novel insights specific to process and measurement model, and making use of them in the IEKF framework, give rise to the real time method with drift-free performance in restricted environments. Furthermore, uncertainty measures, calculated through the method, are valuable information for fusion purposes and also an accurate motion model, derived in this method, can be used as a robust and an accurate localization procedure in different structured environments. These issues are validated through experimental implementations; experiments verify method's efficiency both in pure localization and in SLAM scenarios in the restricted environments, involving loop closures.

This paper introduces an accurate and high speed pose tracking method for mobile robots based on matching of extracted features from consecutive scans. The feature extraction algorithm proposed in this paper uses a global information of the whole scan data and local information around feature points. Uncertainty of each feature is represented using covariance matrices determined due to observation and quantization error. Taking into account each feature's uncertainty in pose shift calculation leads to an accurate estimation of robot pose. Experiments with low range URG_X002 laser range scanner illustrate the effectiveness of the proposed method for mobile robot localization.

in this paper, we propose a new technique in wavelet domain for real time object detection and tracking in a sequence of images. The object to be tracked is identified in the first frame. Our proposed algorithm consists of two phases: the first, wavelet based edge detection is used to form ground boundary map. Then, Object dimensions estimation is implemented to determine probabilistic object areas. finally, target detection based on finding best match using feature vectors is applied. We defined dispersion of wavelet detail coefficient in object area as feature to be matched. Also we proposed a new color model for images to be used in processing algorithm. Our experimental results show that the algorithm is robust and fast. It is also insensitive to changing illumination condition and size of target