Kamil Wereszczyński

Identifying a Joint in Medical Ultrasound Images Using Trained Classifiers

A novel learning approach for detecting the joint in ultrasound images is proposed as a first step of an automated method of assessment of synovitis activity. The training and test data sets consist of images with labeled pixels of the joint region. Feature descriptors based on a pixel’s neighborhood, are selected among SURF, SIFT, FAST, ORB, BRISK, FREAK descriptors, and their mixtures, to define the feature vectors for a trainable pixel classifier. Multiple pixel classifiers, including k-nearest neighbor, support vector machine, and decision tree classifier, are constructed by supervised learning. The AUC measure computed from ROC curves is used as the performance criterion for evaluation. The measure is used to compare and select the best mixture of image descriptors, forming a feature vector for the classifier, the best classifier and the best chain of image preprocessing operations. The final joint detector is a result of clustering the pixels classified as ”joint”. The results of experiments using the proposed method on a set of ultrasound images are presented, demonstrating the method’s applicability and usefulness.

Optimization of Joint Detector for Ultrasound Images Using Mixtures of Image Feature Descriptors

Joint detector is an essential part of an approach towards automated assessment of synovitis activity, which is a subject of the current research work. A recent formulation of the joint detector, that integrates image processing, local image neighborhood descriptors, such as SURF, FAST, ORB, BRISK, FREAK, trainable classification (SVM, NN, CART) and clusterization, results in a large number of possible choices of classifiers, their modes, components of features vectors, and parameter values, and making such choices by experimentation is impractical. This article presents a novel approach, and an implemented environment for the parameter selection process for the joint detector, which automatically choses the best configuration of image processing operators, type of image neighborhood descriptors, the form of a classifier and the clustering method and their parameters. Its implementation uses new scripting tools and generic techniques, such as chain-of-responsibility design pattern and metafunction idiom. Also presented are novel results, comparing the effect of feature vectors composed from multiple SURF descriptors on the performance of the joint detector, which demonstrate the potential of mixture of descriptors for improving the classification results.

VMASS: Massive Dataset of Multi-camera Video for Learning, Classification and Recognition of Human Actions

Expansion of capabilities of intelligent surveillance systems and research in human motion analysis requires massive amounts of video data for training of learning methods and classifiers and for testing the solutions under realistic conditions. While there are many publicly available video sequences which are meant for training and testing, the existing video datasets are not adequate for real world problems, due to low realism of scenes and acted out human behaviors, relatively small sizes of datasets, low resolution and sometimes low quality of video.

Recent Developments in Tracking Objects in a Video Sequence

Methods of tracking of multiple objects or people in video sequences have applications in many fields such as surveillance, art, transport or biology. This, over four decades old area is still very active, with multiple new contributions presented every year. Tracking methods must solve intricate problems, for example occlusion of many objects, crowded scenes, illumination of different places and motion of camera. This paper presents a brief survey of recent developments in video tracking based methods, focused mainly on the last three years. The surveyed methods are divided into two groups: tracking by detection, which includes methods that solve the problem of time-linking objects detected in all video frames, and tracking by correlation, containing methods that follow a selected object using cross correlation. The reviewed methods are collected in a table that lists for each method the benchmark datasets used for its evaluation, implementation environment, and whether it can track single or multiple objects.

Selected Space-Time Based Methods for Action Recognition

A survey on very recent and efficient space-time methods for action recognition is presented. We select the methods with highest accuracy achieved on the challenging datasets such as: HMDB51, UCF101 and Hollywood2. This research focuses on two main space-time based approaches, namely the hand-crafted and deep learning features. We intuitively explain the selected pipelines and review good practices used in state-of-the-art methods including the best descriptors, encoding methods, deep architectures and classifiers. The best methods were chosen and some of them were explained in more details. Furthermore, we conclude how to improve the methods in speed as well as in accuracy and propose directions for further work.

Registration of Ultrasound Images for Automated Assessment of Synovitis Activity

Ultrasound images of joints are used by doctors to assess a degree of synovitis activity, in diagnosis and treatment of rheumatoid arthritis. Research on automation of synovitis assessment from ultrasound images is being conducted, with objectives of lowering medical costs and improving patients care. Analysis of synovitis area in an image should be done relative to the joint and bones, therefore the joint and bones must be located in the initial step. An approach is proposed for locating joint and bones, by registering structural descriptions of the joint region. A preliminary result is presented that includes a description of a registration method that iteratively improves the registration quality, and its application example based on synthetic data.

Camera Calibration and Navigation in Networks of Rotating Cameras

Camera calibration is one of the basic problems concerning intelligent video analysis in networks of multiple cameras with changeable pan and tilt (PT). Traditional calibration methods give satisfactory results, but are human labour intensive. In this paper we introduce a method of camera calibration and navigation based on continuous tracking, which requires minimal human involvement. After the initial pre-calibration, it allows the camera pose to be calculated recursively in real time on the basis of the current and previous camera images and the previous pose. The method is suitable if multiple coplanar points are shared between views from neighbouring cameras, which is often the case in the video surveillance systems.

Optical Flow Based Face Anonymization in Video Sequences.

In this paper we present a method of anonymization of people’s faces in video. Results are analyzed on the basis of optical flow methods. Anonymization bases on face detection. Because of mistakes made by such detectors in video sequences, gaps and false detections appear. They are recognized using the results of face detections and optical flow analysis. In this paper we describe: face detectors and the results of method of analysis of optical-flow based detector. We present novel method of filing gaps and false detection recognizing with use of optical flow. Then we present visual results.

Evaluation of human dynamic balance in Grassmann manifold

The authors present an application of Grassmann manifold to the evaluation of human dynamic balance based on the time series representing movements of hip, knee and ankle joints in the sagittal, frontal and transverse planes. Time series were extracted from gait sequences which were recorded in the Human Motion Laboratory (HML) of the Polish-Japanese Academy of Information Technology in Bytom, Poland using the Vicon system.

Recent Developments on 2D Pose Estimation From Monocular Images

Human pose estimation from monocular images is one of the most significant aspects of modern computer vision tasks and its application demand is still increasing in such areas as automatic images indexing or human activity recognition from video. Among many approaches applied in these areas the one based on pose estimation gives, beyond all doubts, one of the most powerful representation of human on the picture in sense of sparsity and semantics. In this paper we provide a detailed survey of the most efficient methods in 2D pose estimation domain as well as the test results of selected methods on the LSP dataset, which is commonly used by state-of-the-art works.