Roman Juránek

Fully Automatic Roadside Camera Calibration for Traffic Surveillance

This paper deals with automatic calibration of roadside surveillance cameras. We focus on parameters necessary for measurements in traffic-surveillance applications. Contrary to the existing solutions, our approach requires no a priori knowledge, and it works with a very wide variety of road settings (number of lanes, occlusion, quality of ground marking), as well as with practically unlimited viewing angles. The main contribution is that our solution works fully automatically-without any percamera or per-video manual settings or input whatsoever-and it is computationally inexpensive. Our approach uses tracking of local feature points and analyzes the trajectories in a manner based on cascaded Hough transform and parallel coordinates. An important assumption for the vehicle movement is that at least a part of the vehicle motion is approximately straight-we discuss the impact of this assumption on the applicability of our approach and show experimentally that this assumption does not limit the usability of our approach severely. We efficiently and robustly detect vanishing points, which define the ground plane and vehicle movement, except for the scene scale. Our algorithm also computes parameters for radial distortion compensation. Experiments show that the obtained camera parameters allow for measurements of relative lengths (and potentially speed) with ~2% mean accuracy. The processing is performed easily in real time, and typically, a 2-min-long video is sufficient for stable calibration.

Real-Time Pose Estimation Piggybacked on Object Detection

We present an object detector coupled with pose estimation directly in a single compact and simple model, where the detector shares extracted image features with the pose estimator. The output of the classification of each candidate window consists of both object score and likelihood map of poses. This extension introduces negligible overhead during detection so that the detector is still capable of real time operation. We evaluated the proposed approach on the problem of vehicle detection. We used existing datasets with viewpoint/pose annotation (WCVP, 3D objects, KITTI). Besides that, we collected a new traffic surveillance dataset COD20k which fills certain gaps of the existing datasets and we make it public. The experimental results show that the proposed approach is comparable with state-of-the-art approaches in terms of accuracy, but it is considerably faster - easily operating in real time (Matlab with C++ code). The source codes and the collected COD20k dataset are made public along with the paper.

Local Rank Differences Image Feature Implemented on GPU

A currently popular trend in object detection and pattern recognition is usage of statistical classifiers, namely AdaBoost and its modifications. The speed performance of these classifiers largely depends on the low level image features they are using: both on the amount of information the feature provides and the executional time of its evaluation. Local Rank Differences is an image feature that is alternative to commonly used haar wavelets. It is suitable for implementation in programmable (FPGA) or specialized (ASIC) hardware, but --- as this paper shows --- it performs very well on graphics hardware (GPU) as well. The paper discusses the LRD features and their properties, describes an experimental implementation of LRD in graphics hardware, presents its empirical performance measures compared to alternative approaches and suggests several notes on practical usage of LRD and proposes directions for future work.

High performance architecture for object detection in streamed videos

In this paper, we introduce a novel architecture of an engine for high performance multi-scale detection of objects in videos based on WaldBoost training algorithm. The key properties of the architecture include processing of streamed data and low resource consumption. We implemented the engine in FPGA and we show that it can process 640×480 pixel video streams at over 160 fps without the need of external memory. We evaluate the design on the face detection task, compare it to state of the art designs, and discuss its features and limitations.

Fast bilateral filter for HDR imaging

Bilateral filtering is a method often used in image processing applications. It is specifically useful for HDR algorithms. A novel approach to a fast and close approximation of bilateral filtering is presented. The method is designed especially with a focus on HDR image conversion into a normal color space processing. This paper presents the methods itself, describes the sources of acceleration and discusses the results of the method.

Low-Level Image Features for Real-Time Object Detection

Object detection in still images and in video sequences has a wide range of applications and while it is a very costly task from the computational resources point of view, very high demand exists for efficient object detection methods and implementations. One of the frequently used techniques of fast object detection is usage of classifiers to scan the image and attempt classification of every potential object position or even every potential position in the image being searched. The classifiers can be implemented as statistical classifiers based on supervised machine learning and can take as their input low-level features (sometimes called weak classifiers) extracted from the window being classified. In principle, such features can be immediately the image pixels, but by using more complex feature extractors, the classifiers can achieve better performance – both in the detection rate and in the speed. This chapter describes several image feature extractors used in real-time object detection and in detail discusses the novel features based on local ranks. The features have been designed so that they have equal descriptive and generalization power as their state-of-theart alternatives, but at the same time to be efficiently implementable in hardware. These features prove to be efficient, not only in the hardware implementations (tested in FPGA chips), but also when implemented using the SSE instruction set of the contemporary CPU’s and implemented in the graphics processors (GPU’s). The classification background and specification of requirements on the low-level image feature extractors is given in section 2. Formal definition of the features based on local ranks is given in section 3. The performance of the LRP image feature extractors was evaluated from the point of view of the classifier construction and the results are also given in section 3. Section 4 describes efficient implementations of the feature extractor on different hardware platforms, namely the SSE instruction set; FPGA (Field-Programmable Gate Arrays) defined in the hardware definition language VHDL; and GPU implementations, using both the shading language GLSL and the CUDA programming environment. Section 4 also contains performance evaluation of the different implementations of the low-level feature extractors. 6

Implementation of the "Local Rank Differences" Image Feature Using SIMD Instructions of CPU

Usage of statistical classifiers, namely AdaBoost and its modifications, in object detection and pattern recognition is a contemporary and popular trend. The computatiponal performance of these classifiers largely depends on low level image features they are using: both from the point of view of the amount of information the feature provides and the executional time of its evaluation. Local rank difference is an image feature that is alternative to commonly used Haar features. It is suitable for implementation in programmable (FPGA) or specialized (ASIC) hardware as well as graphics hardware (GPU). Additionally, as shown in this paper, it performs very well on common CPUpsilas. The paper discusses the LRD features and their properties, describes an experimental implementation of LRD using the multimedia instruction set of current general-purpose processors, presents its empirical performance measures compared to alternative approaches, and suggests several notes on practical usage of LRD and proposes directions for future work.

Analysis of wear debris through classification

This paper introduces a novel method of wear debris analysis through classification of the particles based on machine learning. Wear debris consists of particles of metal found in e.g. lubricant oils used in engineering equipment. Analytical ferrography is one of methods for wear debris analysis and it is very important for early detection or even prevention of failures in engineering equipment, such as combustion engines, gearboxes, etc. The proposed novel method relies on classification of wear debris particles into several classes defined by the origin of such particles. Unlike the earlier methods, the proposed classification approach is based on visual similarity of the particles and supervised machine learning. The paper describes the method itself, demonstrates its experimental results, and draws conclusions.

Comparison of bubble detectors and size distribution estimators

Abstract Detection, counting and characterization of bubbles, that is, transparent objects in a liquid, is important in many industrial applications. These applications include monitoring of pulp delignification and multiphase dispersion processes common in the chemical, pharmaceutical, and food industries. Typically the aim is to measure the bubble size distribution. In this paper, we present a comprehensive comparison of bubble detection methods for challenging industrial image data. Moreover, we compare the detection-based methods to a direct bubble size distribution estimation method that does not require the detection of individual bubbles. The experiments showed that the approach based on a convolutional neural network (CNN) outperforms the other methods in detection accuracy. However, the boosting-based approaches were remarkably faster to compute. The power spectrum approach for direct bubble size distribution estimation produced accurate distributions and it is fast to compute, but it does not provide the spatial locations of the bubbles. Selecting the most suitable method depends on the specific application.

Holistic recognition of low quality license plates by CNN using track annotated data

This work is focused on recognition of license plates in low resolution and low quality images. We present a methodology for collection of real world (non-synthetic) dataset of low quality license plate images with ground truth transcriptions. Our approach to the license plate recognition is based on a Convolutional Neural Network which holistically processes the whole image, avoiding segmentation of the license plate characters. Evaluation results on multiple datasets show that our method significantly outperforms other free and commercial solutions to license plate recognition on the low quality data. To enable further research of low quality license plate recognition, we make the datasets publicly available.