Search

Object reidentification (ReID) in cluttered rigid scenes is a challenging problem especially when same-looking objects coexist in the scene. ReID is accepted to be one of the most powerful tools for matching the correct identities to each individual object when issues such as occlusion, missed detections, multiple same-looking objects coexisting in the same scene, and disappearance of objects from the view and/or revisiting the same region arise. We propose a novel framework towards more efficient object ReID, improved object reidentification (IO-ReID), to perform object ReID in challenging scenes with real-time processing in mind. The proposed approach achieves distinctive and efficient object embedding via training with the triplet loss, with input from both the foreground/background split by bounding box, and the full input image. With extensive experiments on two datasets serving for Object ReID, we demonstrate that the proposed method, IO-ReID, obtains a higher ReID accuracy and runs faster compared to the state-of-the-art methods on object ReID. [ABSTRACT FROM AUTHOR]

Re-identifying objects in a rigid scene across varying viewpoints (object Re-ID) is a challenging task, in particular when there are similar, even identical objects coexist in the same environment. Discriminative features play no doubt an essential role in addressing this challenge, while for practical deployment, real-time performance is another desired attribute. We therefore propose a novel framework, named Fast re-OBJ, that is able to improve both Re-ID accuracy and processing speed via tight coupling between the instance segmentation module and embedding generation module. The rich object encoding in the instance segmentation backbone is directly shared to the embedding generation module for training a more discriminative representation via a triplet network. Moreover, we create datasets with the segmentation outputs using real-time object detectors to train and evaluate our object embedding module. With extensive experiments, we prove that our proposed Fast re-OBJ improves the object Re-ID accuracy by 5% and the speed is 5 × faster compared to the state-of-the-art methods. The dataset and code repository are publicly available at: . [ABSTRACT FROM AUTHOR]

The process of object tracking involves consistently identifying each instance across frames depending on initial set of object detection(s). Moreover, in multiple object tracking (MOT), the process through tracking-by-detection paradigm consists of performing two common steps consecutively, which are detection and data association. In MOT, it is targeted to associate detections across frames by localizing and identifying all objects of interest. MOT algorithms further keep tracking even the most challenging issues such as revisiting the same view, missing detections, occlusion and temporarily unseen objects, same-appearance objects coexisting in the same frame occur. Hence, re-identification (re-id) appears to be the most powerful tool for assigning the correct identities to each individual instance when aforementioned issues arise. In this work, we propose a similarity-based person re-id framework, called SAT, using a Siamese neural network via shared weights. Once detections are obtained from the backbone SAT applies a Siamese feature extraction model and then we introduce a similarity array for assessing tracklet(s) and detection(s). We examine the performance of SAT on several benchmarks with extensive experiments and statistical tests, where we improve the current state-of-the-art according to commonly used performance metrics with higher accuracy, less ID switches, less false positive and negative rates. [ABSTRACT FROM AUTHOR]

Recent use scenarios involving human–robot collaboration have revealed that the robots require elastic joints to safely interact with humans. It is also critical to know applied force/torque (f/t) during the interaction for control and motion planning purposes. In this article, we estimate the external f/t values without using any sensors other than low-cost encoders by exploiting the inherent elastic properties of the joint. For estimation, the following two different approaches are used: model based and model free. In the model-based approach, an extended Kalman filter (EKF) and an external force observer (EFOB) are used considering the dynamical behavior of the system to estimate the interaction force. In the model-free approach, the artificial neural network (ANN) utilizes the data gathered from mechanical systems. In comparative analysis, we have, therefore, considered three different estimation methods, two of which are model based and the remaining one is model free (i.e., data driven). Implementing these estimation algorithms experimentally on a variable stiffness joint, we performed an extensive evaluation of their performances. All methods show similar level of performance in terms of the root-mean-square (RMS) error with 0.0847, 0.0841, and 0.1082 N for the EKF, EFOB, and ANN, respectively. Model-based methods do not require continuous data stream through the experimental set up. On the other hand, the ANN does not need an explicit model of the system; therefore, it may become preferable when the detailed model derivation is not possible. [ABSTRACT FROM AUTHOR]

The progress in technology has led to an increase in the amount of data that can be generated and processed. Downsampling methods are employed to eliminate unnecessary data but can also lead to the loss of valuable information. Herein, we designed a pooling algorithm, Conditional-Pooling , that provides a transitive structure composed of average (avg.) and max-pooling methods, and it hosts the advantageous behaviors from both. We examined our approach on several image recognition and detection tasks using deep neural networks. The results reveal that our method excels over major rivals like max-pooling and avg. pooling when tested on MNIST, Fashion-MNIST, and CIFAR for classification, Pascal VOC for detection, COCO minitrains, and ADORESet for both classification and detection. The output of Conditional-Pooling preserves important image features such as edges and corners and contains more salient features that lead to more accurate results. The code for Conditional-Pooling will be available at https://github.com/bayraktare/conditional_pooling. • Adaptive pooling for novel precision and accuracy especially for complex scenes. • Exceeding existing methods on popular visual tasks regarding various metrics. • Game-changing results for noise elimination and image downsampling. [ABSTRACT FROM AUTHOR]

Object detection still keeps its role as one of the fundamental challenges within the computer vision territory. In particular, achieving satisfying results concerning object detection from outdoor images occupies a considerable space. In this study, in addition to comparing handcrafted feature detector/descriptor performance with deep learning methods over ornamental plant images at the outdoor, we propose a framework to improve the detection of these plants. Firstly, we take query images in the RGB format from the onboard UAV camera. Secondly, our model classifies the scene as a planting or an urban area. Thirdly, if the images are from planting area, thirdly, we filter the field according to the color and acquire only the green parts. Lastly, we feed the object detector model with the filtered area and obtain the category and localization of the plants as a result. In parallel, we also estimate the number of interested plants using the geometrical relations and predefined average plant size, then we verify the outputs of the object detector with this results. The conducted experiments show that deep learning based object detection methods overtake conventional feature detector/descriptor techniques in terms of accuracy, recall, precision, and sensitivity rates. The field classifier model, VGGNet, achieves a 98.17 % accuracy for this task, whilst YoloV3 achieves 91.6 % accuracy with 0.12 IOU for object detection as the best method. The proposed framework also improves the overall performance of these algorithms by 1.27 % for accuracy and 0.023 for IOU. By specifying the limits thoroughly and developing task-dependent approaches, we reveal the great potential of our framework plant detection and counting in the wild consisting of basic image preprocessing techniques, geometrical operations, and deep neural network. [ABSTRACT FROM AUTHOR]

In recent years, the computer vision applications in the robotics have been improved to approach human-like visual perception and scene/context understanding. Following this aspiration, in this study, we explored the possibility of better object manipulation performance by connecting the visual recognition of objects to their physical attributes, such as weight and center of gravity (CoG). To develop and test this idea, an object manipulation platform is built comprising a robotic arm, a depth camera fixed at the top center of the workspace, embedded encoders in the robotic arm mechanism, and microcontrollers for position and force control. Since both the visual recognition and force estimation algorithms use deep learning principles, the test set-up was named as Deep-Table. The objects in the manipulation tests are selected from everyday life and are common to be seen on modern office desktops. The visual object localization and recognition processes are performed from two distinct branches by deep convolutional neural network architectures. We present five of the possible cases, having different levels of information availability on the object weight and CoG in the experiments. The results confirm that using our algorithm, the robotic arm can move different types of objects successfully varying from several grams (empty bottle) to around 250 g (ceramic cup) without failure or tipping. The proposed method also shows that connecting the object recognition with load estimation and contact point further improves the performance characterized by a smoother motion. [ABSTRACT FROM AUTHOR]

The primary motivation of computer vision in the robotics field is to obtain a perception level that is as close as possible to human visual system. To achieve this, the inclusion of large datasets is necessary, sometimes involving less-frequent and seemingly irrelevant data to increase the system robustness. To minimize the effort and time in forming such extensive datasets from real world, the preferred method is to utilize simulation environments, replicating real-world conditions as much as possible. Following this solution path, the machine vision problems in robotics (i.e., object detection, recognition, and manipulation) often employ synthetic images in datasets and, however, do not mix them with real-world images. When the systems are trained only using the synthetic images and tested within the simulated world, the tasks requiring object recognition in robotics can be accomplished. However, the systems trained using this procedure cannot be directly used in the real-world experiments or end-user products due to the inconsistencies between real and simulation environments. Therefore, we propose a hybrid image dataset including annotated desktop objects from real and synthetic worlds (ADORESet). This hybrid dataset provides purposeful object categories with a sufficient number of real and synthetic images. ADORESet is composed of colored images with the dimension of 300×300 pixels within 30 categories. Each class has 2500 real-world images acquired from the wild web and 750 synthetic images that are generated within Gazebo simulation environment. This hybrid dataset enables researchers to implement their own algorithms for both real-world and simulation environment conditions. ADORESet is composed of fully annotated object images. The limits of objects are manually specified, and the bounding box coordinates are provided. The successor objects are also labeled to give statistical information and the likelihood about the relations of the objects within the dataset. To further demonstrate the benefits of this dataset, it is tested in object recognition tasks by fine-tuning the state-of-the-art deep convolutional neural networks such as VGGNet, InceptionV3, ResNet, and Xception. The possible combinations regarding the data types for these models are compared in terms of time, accuracy, and loss values. As a result of the conducted object recognition experiments, training with all-real images yields approximately 49% validation accuracy for simulation images. When the training is performed with all-synthetic images and validated using all-real images, the accuracy becomes lower than 10%. If the complete ADORESet is employed for training and validation, the hybrid dataset validation accuracy reaches approximately to 95%. This result proves further that including the real and synthetic images together in the training and validation sessions increases the overall system accuracy and reliability. [ABSTRACT FROM AUTHOR]

Highlights • Novel synthesizing methods are proposed for translational VRPs. • The proposed methods are given algorithmically to ease the implementation. • Synthesizing methods are experimentally validated. • A low-cost variable stiffness joint design is presented via VRPs. • Performance of the VSJ with a simple controller is evaluated experimentally. Abstract Robot joints are expected to be safe, compliant, compact, simple and low-cost. Gravity compensation, zero backlash, energy efficiency and stiffness adjustability are some desired features in the robotic joints. The variable radius pulleys (VRPs) provide a simple, compact and low-cost solution to the stiffness adjustment problem. VRP mechanisms maintain a preconfigured nonlinear force-elongation curve utilizing off-the-shelf torsional spring and pulley profile. In this paper, three synthesis algorithms are presented for VRP mechanisms to obtain desired force-elongation curve. In addition, a feasibility condition is proposed to determine the torsional spring coefficient. Using the synthesis methods and the feasibility condition, a variable stiffness mechanism is designed and manufactured which uses two VRPs in an antagonistic cable driven structure. Afterwards, the outputs of three synthesis methods are compared to force-elongation characteristics in the tensile testing experiment. A custom testbed is manufactured to measure the pulley rotation, cable elongation and tensile force at the same time. Using the experiment as the baseline, the best algorithm achieved to reproduce the desired curve with a root-mean-square (RMS) error of 13.3%. Furthermore, VRP-VSJ is implemented with a linear controller to reveal the performance of the mechanism in terms of position accuracy and stiffness adjustability. [ABSTRACT FROM AUTHOR]

In digital era, being stranded from very basic telecommunication protocols and internet makes vehicle rerouting-like crucial tools more difficult or even impossible, especially in times of disaster and emergency. In this study, we propose a modular rerouting framework for only one single vehicle composed of visual perception, property estimation and trajectory optimization, which enables to generate optimum paths exploiting aerial imagery. Once the deep network, which is fine-tuned on newly introduced dataset herein named MaVefAI, processes the input, the following module estimates pose, motion direction and speed of detected vehicles. Afterwards, we link the appropriate vehicles via graphs to obtain group properties that pave the way for estimating the traffic congestion level. In the end, we get the output as the optimum path from the independent trajectory optimization module to which required inputs are already sent by preceding modules. We solve the multi-objective cost function subject to velocity and congestion intervals, which comprises distance, traffic congestion level, and angle inconsistency. We employ Dijkstra, A*, RRT, and RRT* to optimize the cost while vast majority of existing works focus to optimize via single method. The fine-tuned segmentation model accuracy becomes more than 98% for vehicle groups thanks to MaVefAI. The extensive experiments reveal that all algorithms follow the same path. However, RRT* achieves the fastest result by examining most of the possible options in less time, which also appears to be the most robust method comparing with the alternatives for route optimization. Our dataset MaVefAI is publicly available here: https://precisioncomputing.sakarya.edu.tr. [ABSTRACT FROM AUTHOR]

Optimal path planning is a fundamental problem in artificial intelligence (AI) and has wide applications in areas such as robotics, transportation, and logistics. In this paper, we propose a novel approach for ordering algorithms in pathfinding problems using the concept of shell layers. Our approach aims to improve the computational efficiency and distance cost of ordering algorithms. To evaluate the effectiveness of our approach, we compare it with five state-of-the-art techniques commonly used in the field of ordering algorithms on eight different scenarios with varying configurations. Our results show that our proposed approach outperformed the state-of-the-art techniques in terms of time-computational complexity and distance cost in most of the scenarios, demonstrating its potential as a new state-of-the-art technique for ordering algorithms. Specifically, our approach outperformed the nearest neighbor, A*, branch and bound, Christofides, and genetics ordering algorithms. However, we also identified two specific situations where our approach was outperformed by two state-of-the-art algorithms due to the specific distribution of the goal nodes. These findings highlight the importance of evaluating and comparing ordering algorithms in various scenarios. Our approach has significant implications for AI research and development, as it has the potential to improve the performance of pathfinding algorithms in various applications. We also discuss the limitations of our approach and potential areas for further research, such as investigating the effectiveness of our approach in different types of graphs and exploring the potential use of machine learning techniques to optimize the shell layer construction. The code for this study is available at https://github.com/abdullah1aloush1/MuGONA. • Equally-spaced nodes formed over workspace with obstacles, optimizing search process. • Efficient neighboring graph built for equally-spaced nodes, reducing complexity. • Path optimization for navigation by ordering goal nodes, reducing time and distance. • Path smoothing for navigation via interpolation to yield a trajectory on path nodes. [ABSTRACT FROM AUTHOR]