Search

This paper proposes a method to enhance video object detection for indoor environments in robotics. Concretely, it exploits knowledge about the camera motion between frames to propagate previously detected objects to successive frames. The proposal is rooted in the concepts of planar homography to propose regions of interest where to find objects, and recursive Bayesian filtering to integrate observations over time. The proposal is evaluated on six virtual, indoor environments, accounting for the detection of nine object classes over a total of ∼ 7k frames. Results show that our proposal improves the recall and the F1-score by a factor of 1.41 and 1.27, respectively, as well as it achieves a significant reduction of the object categorization entropy (58.8%) when compared to a two-stage video object detection method used as baseline, at the cost of small time overheads (120 ms) and precision loss (0.92).

Efficient yet accurate extraction of depth from stereo image pairs is required by systems with low power resources, such as robotics and embedded systems. State-of-the-art stereo matching methods based on convolutional neural networks require intensive computations on GPUs and are difficult to deploy on embedded systems. In this paper, we propose MTStereo2.0, an improved version of the MTStereo stereo matching method, which includes a more robust context-driven cost function, better detection of incorrect matches and the computation of disparity at pixel level. MTStereo provides accurate sparse and semi-dense depth estimation and does not require intensive GPU computations. We tested it on several benchmark data sets, namely KITTI 2015, Driving, FlyingThings3D, Middlebury 2014, Monkaa and the TrimBot2020 garden data sets, and achieved competitive accuracy. The code is available at https://github.com/rbrandt1/MaxTreeS.

Monitoring of public and private places is of great importance for security of people and is usually done by means of surveillance cameras. In this paper we propose an approach for monitoring of roads, to detect car crashes and tire skidding, based on the analysis of sound signals, which can complement or, in some cases, substitute video analytic systems. The system that we propose employs a MobileNet deep architecture, designed to efficiently run on embedded appliances and be deployed on distributed systems for road monitoring. We designed a recognition system based on analysis of audio frames and tested it on the publicly available MIVIA road events data set. The performance results that we achieved (recognition rate higher than 99%) are higher than existing methods, demonstrating that the proposed approach can be deployed on embedded devices in a distributed surveillance system.

This paper describes preliminary work in the recent promising approach of generating synthetic training data for facilitating the learning procedure of deep learning (DL) models, with a focus on aerial photos produced by unmanned aerial vehicles (UAV). The general concept and methodology are described, and preliminary results are presented, based on a classification problem of fire identification in forests as well as a counting problem of estimating number of houses in urban areas. The proposed technique constitutes a new possibility for the DL community, especially related to UAV-based imagery analysis, with much potential, promising results, and unexplored ground for further research., Workshop on Deep-learning based computer vision for UAV in conjunction with CAIP 2019, Salerno, italy, September 2019

The monitoring of farm animals is important as it allows farmers keeping track of the performance indicators and any signs of health issues, which is useful to improve the production of milk, meat, eggs and others. In Europe, bovine identification is mostly dependent upon the electronic ID/RFID ear tags, as opposed to branding and tattooing. The RFID based ear-tagging approach has been called into question because of implementation and management costs, physical damage and animal welfare concerns. In this paper, we conduct a case study for individual identification of Holstein cattle, characterized by black, brown and white patterns, in collaboration with the Dairy campus in Leeuwarden. We use a FLIR E6 thermal camera to collect an infrared and RGB image of the side view of each cow just after leaving the milking station. We apply a fully automatic pipeline, which consists of image processing, computer vision and machine learning techniques on a data set containing 1237 images and 136 classes (i.e. individual animals). In particular, we use the thermal images to segment the cattle from the background and remove horizontal and vertical pipes that occlude the cattle in the station, followed by filling the blank areas with an inpainting algorithm. We use the segmented image and apply transfer learning to a pre-trained AlexNet convolutional neural network. We apply five-fold cross-validation and achieve an average accuracy rate of 0.9754 ± 0.0097. The results obtained suggest that the proposed non-invasive approach is highly effective in the automatic recognition of Holstein cattle from the side view. In principle, this approach is applicable to any farm animals that are characterized by distinctive coat patterns.

Visual place recognition is an important component of systems for camera localization and loop closure detection. It concerns the recognition of a previously visited place based on visual cues only. Although it is a widely studied problem for indoor and urban environments, the recent use of robots for automation of agricultural and gardening tasks has created new problems, due to the challenging appearance of garden-like environments. Garden scenes predominantly contain green colors, as well as repetitive patterns and textures. The lack of available data recorded in gardens and natural environments makes the improvement of visual localization algorithms difficult. In this paper we propose an extended version of the TB-Places data set, which is designed for testing algorithms for visual place recognition. It contains images with ground truth camera pose recorded in real gardens in different seasons, with varying light conditions. We constructed and released a ground truth for all possible pairs of images, indicating whether they depict the same place or not. We present the results of a benchmark analysis of methods based on convolutional neural networks for holistic image description and place recognition. We train existing networks (i.e. ResNet, DenseNet and VGG NetVLAD) as backbone of a two-way architecture with a contrastive loss function. The results that we obtained demonstrate that learning garden-tailored representations contribute to an improvement of performance, although the generalization capabilities are limited.

Nowadays, much is known about the functions of the components of the human auditory system. Computational models of these components are widely accepted and recently inspired the work of researchers in pattern recognition and signal processing. In this work we present a novel filter, which we call COPE (Combination of Peaks of Energy), that is inspired by the way the sound waves are converted into neuronal firing activity on the auditory nerve. A COPE filter creates a model of the pattern of the neural activity generated by a sound of interest and is able to detect the same pattern and modified versions of it. We apply the proposed method on the task of event detection for surveillance of roads. For the experiments, we use a publicly available data set, namely the MIVIA road events data set. The results that we achieve (recognition rate equal to \(94\%\) and false positive rate lower than \(4\%\)) and the comparison with existing methods demonstrate the effectiveness of the proposed bio-inspired filters for audio analysis.

Glaucoma is the second leading cause of irreversible vision loss. Early diagnosis and treatment can, however, slow the progression of the disease. Specialists making this diagnosis rely on several tests and examinations such as visual field tests and examinations of retinal images and optical coherence tomography images. One of the regions examined by specialists when checking for retinal conditions is the optic nerve head region, which is the brightest region in retinal images. Within this region, the ratio between the cup and the disc can be used when diagnosing for glaucoma. Calculating the cup–disc ratio requires the segmentation of both the disc and the cup from retinal images. In a previous paper, a method for segmenting the disc was proposed. Here another deep learning model, H-OCS, is proposed for segmenting the cup from retinal images. A customized InceptionV3 model with transfer learning and image augmentation is used. Additionally, the output of H-OCS is refined and enhanced using a series of post-processing steps. H-OCS is tested on six publicly available datasets: RimOneV3, Drishti, Messidor, Refuge, Riga, and Magrebia and several ablation studies are conducted to evaluate the effectiveness of the proposed approach. Additionally, the performance of H-OCS is compare with other studies. An overall average accuracy of 97.86%, DC of 88.37%, Sensitivity of 89.09% and IoU of 79.66% was achieved.

We propose a novel means to improve the accuracy of semantic segmentation based on multi-task learning. More specifically, in our Multi-Task Semantic Segmentation and Super-Resolution (MT-SSSR) framework, we jointly train a super-resolution and semantic segmentation model in an end-to-end manner using the same task loss for both models. This allows us to optimize the super-resolution model towards producing images that are optimal for the segmentation task, rather than ones that are of high-fidelity. Simultaneously we adapt the segmentation model to better utilize the improved images and thereby improve the segmentation accuracy. We evaluate our approach on multiple public benchmark datasets, and our extensive experimental results show that our novel MT-SSSR framework outperforms other state-of-the-art approaches.

Fast, non-destructive and on-site quality control tools, mainly high sensitive imaging techniques, are important to assess the reliability of photovoltaic plants. To minimize the risk of further damages and electrical yield losses, electroluminescence (EL) imaging is used to detect local defects in an early stage, which might cause future electric losses. For an automated defect recognition on EL measurements, a robust detection and rectification of modules, as well as an optional segmentation into cells is required. This paper introduces a method to detect solar modules and crossing points between solar cells in EL images. We only require 1-D image statistics for the detection, resulting in an approach that is computationally efficient. In addition, the method is able to detect the modules under perspective distortion and in scenarios, where multiple modules are visible in the image. We compare our method to the state of the art and show that it is superior in presence of perspective distortion while the performance on images, where the module is roughly coplanar to the detector, is similar to the reference method. Finally, we show that we greatly improve in terms of computational time in comparison to the reference method.

In this paper, we propose a novel method to adaptively select the most informative and least redundant feature subset, which has strong discriminating power with respect to the target label. Unlike most traditional methods using vectorial features, our proposed approach is based on graph-based features and thus incorporates the relationships between feature samples into the feature selection process. To efficiently encapsulate the main characteristics of the graph-based features, we probe each graph structure using the steady state random walk and compute a probability distribution of the walk visiting the vertices. Furthermore, we propose a new information theoretic criterion to measure the joint relevance of different pairwise feature combinations with respect to the target feature, through the Jensen-Shannon divergence measure between the probability distributions from the random walk on different graphs. By solving a quadratic programming problem, we use the new measure to automatically locate the subset of the most informative features, that have both low redundancy and strong discriminating power. Unlike most existing state-of-the-art feature selection methods, the proposed information theoretic feature selection method can accommodate both continuous and discrete target features. Experiments on the problem of P2P lending platforms in China demonstrate the effectiveness of the proposed method.

We propose a hierarchical architecture for human behavior analysis.Micro-behaviors are analyzed by a time-delay neural network.Macro-behaviors are analyzed by a fuzzy system.The experiments have been performed on the standard CAVIAR datasets. Analysis and detection of human behaviors from video sequences has became recently a very hot research topic in computer vision and artificial intelligence. Indeed, human behavior understanding plays a fundamental role in several innovative application domains such as smart video surveillance, ambient intelligence and content-based video information retrieval. However, the uncertainty and vagueness that typically characterize human daily activities make frameworks for human behavior analysis (HBA) hard to design and develop. In order to bridge this gap, this paper proposes a hierarchical architecture, based on a tracking algorithm, time-delay neural networks and fuzzy inference systems, aimed at improving the performance of current HBA systems in terms of scalability, robustness and effectiveness in behavior detection. Precisely, the joint use of the aforementioned methodologies enables both a quantitative and qualitative behavioral analysis that efficiently face the intrinsic people/objects tracking imprecision and provide context aware and semantic capabilities for better identifying a given activity. The validity and effectiveness of the proposed framework have been verified by using the well-known CAVIAR dataset and comparing our system's performance with other similar approaches working on the same dataset.

The psychological overcharge issue related to human inadequacy to maintain a constant level of attention in simultaneously monitoring multiple visual information sources makes necessary to develop enhanced video surveillance systems that automatically understand human behaviors and identify dangerous situations. This paper introduces a semantic human behavioral analysis (HBA) system based on a neuro-fuzzy approach that, independently from the specific application, translates tracking kinematic data into a collection of semantic labels characterizing the behavior of different actors in a scene in order to appropriately classify the current situation. Different from other HBA approaches, the proposed system shows high level of scalability, robustness and tolerance for tracking imprecision and, for this reason, it could represent a valid choice for improving the performance of current systems. © 2012 IEEE.

Nowadays the measurement of the nuchal translucency thickness is being used as part of routine ultrasound scanning during the end of the first trimester of pregnancy, for the screening of chromosomal defects, as trisomy 21. Currently, the measurement is being performed manually by physicians. The measurement can take a long time for being accomplished, needs to be performed by highly skilled operators, and is prone to errors. In this paper we present an algorithm that automatically detects the border of the nuchal translucency, once a region of interest has been manually identified. The algorithm is based on the minimisation of a cost function, and the optimisation is performed using the dynamic programming paradigm. The method we present overcomes several of the drawbacks present in the state of the art algorithms.

International audience; In a video surveillance system the object tracking is one of the mostchallenging problem. In fact objects in the world exhibit complex interactions.When captured in a video sequence, some interactions manifest themselves asocclusions. A visual tracking system must be able to track objects which arepartially or even fully occluded. In this paper we present a novel method oftracking objects through occlusions using a multi-resolution representation ofthe moving regions. The matching between objects in two consecutive framesto recognize the trajectories is preformed in a graph theoretic approach. The ex-perimental results on the standard database PEST2001 show that the approachlooks promising.