Michael Hödlmoser

Camera auto-calibration using pedestrians and zebra-crossings

In this paper we present a novel camera self-calibration technique to automatically recover intrinsic and extrinsic parameters of a static surveillance camera by observing a traffic scene. The scene must consist of one or more pedestrians and a zebra-crossing. We first extract a horizontal vanishing point and a vanishing line from a zebra-crossing. The observation of pedestrians allows calculating a so called vertical line of mass. All lines of mass are parallel in 3D space and therefore the vertical vanishing point can be estimated. The second horizontal vanishing point can be calculated by introducing the triangle spanned by three orthogonal vanishing points. All three vanishing points are then taken to gather the intrinsic parameters. The extrinsic parameters are calculated after the determination of the cameras height from the distance between two zebra-crossing edges. By combining static and dynamic calibration objects, the method gets robust against outliers. This robustness in combination with the practicability is shown in our experiments, which are carried out by using synthetic and real data of different application scenarios.

Pedestrian detection implemented on a fixed-point parallel architecture

This paper describes the implementation of a real-time pedestrian detector on a single instruction, multiple data (SIMD), fixed-point digital signal processor (DSP). We reformulate the Histogram of Oriented Gradients algorithm for calculation with a relatively simple instruction set architecture (ISA) and partition the image for parallel processing. Results obtained using an ISA simulator indicate a maximum frame rate above 40fps for 1MPixel images, with a detection accuracy comparable to a double-precision floating-point reference implementation.

Multiple camera self-calibration and 3D reconstruction using pedestrians

The analysis of human motion is an important task in various surveillance applications. Getting 3D information through a calibrated system might enhance the benefits of such analysis. This paper presents a novel technique to automatically recover both intrinsic and extrinsic parameters for each surveillance camera within a camera network by only using a walking human. The same feature points of a pedestrian are taken to calculate each cameras intrinsic parameters and to determine the relative orientations of multiple cameras within a network as well as the absolute positions within a common coordinate system. Experimental results, showing the accuracy and the practicability, are presented at the end of the paper.

Classification and Pose Estimation of Vehicles in Videos by 3D Modeling within Discrete-Continuous Optimization

This paper presents a framework for classification and pose estimation of vehicles in videos by assuming their given 3D models. We rank possible poses and types for each frame and exploit temporal coherence between consecutive frames for refinement. As a novelty, first, we cast the estimation of a vehicles pose and type as a solution of a continuous optimization problem over space and time. Due to a non-convexity of this problem, good initial starting points are important. We propose to obtain them by a discrete temporal optimization reaching a global optimum on a ranked discrete set of possible types and poses. Second, to guarantee effectiveness of the proposed discrete-continuous optimization, we present a novel way to efficiently reduce the search space of potential 3D model types and poses for each frame for the discrete optimizer. It avoids common expensive evaluation of all possible discretized hypotheses. The key idea towards efficiency lies in a novel combination of detecting the vehicle, rendering the 3D models, matching projected edges to input images, and using a tree structured Markov Random Field to get fast and globally optimal inference and to force the vehicle follow a feasible motion model in the initial phase. Quantitative and qualitative experiments on a variety of videos with vast variation of vehicle types show superior results to state-of-the-art methods.

Automatic Coin Classification and Identification

We investigate object recognition and classification in a setting with a large number of classes as well as recognition and identification of individual objects of high similarity. Real-world data sets were obtained for the classification and identification tasks. The considered classification task is the discrimination of modern coins into several hundreds of different classes. Identification is investigated for hand-made ancient coins. Intra-class variance due to wear and abrasion vs. small inter-class variance makes the classification of modern coins challenging. For ancient coins the intra-class variance makes the identification task possible, as the appearance of individual hand-struck coins is unique. Figure 1 shows sample images for the considered collections of coins.

3D acquisition of historical coins and its application area in numismatics

Nowadays, 2D photography is the common technique for the documentation and digitalization of historical coin inventories. However, by using 2D photography a huge amount of information is lost due to the projection of a 3D structure onto a 2D image. A solution to this problem would be the use of 3D scanning devices to obtain accurate 3D models of the coins. In this paper we show results of scanning 24 historical coins from the Roman and medieval age using a high-accuracy active stereo scanner. We furthermore highlight the various benefits of this acquisition method for coin documentation, coin measurement and coin recognition. The results show that accurate 3D models can be obtained despite the small size and high reflectance of historical coins.

Model-Based Vehicle Pose Estimation and Tracking in Videos Using Random Forests

This paper presents a computational effective framework for tracking and pose estimation of vehicles in videos reaching comparable performance to state-of-the-art methods. We cast the problem of vehicle tracking as ranking possible poses for each frame and connecting subsequent poses by exploiting a feasible motion model over time. As a novelty, we use random forests trained on a set of existing 3D models for estimating the pose. We discretize the viewpoint space for training, where a synthetic camera is orbiting around the models. To compare projections of 3D models to real world 2D input frames, we introduce simple but discriminative principle gradient features to describe both images. A Markov Random Field ensures to pick the perfect pose over time and the vehicle to follow a feasible motion. As can be seen from our experiments performed on a variety of videos with vast variation of vehicle types, the proposed framework achieves similar results in less computational time compared to state-of-the-art methods.

Surface Layout Estimation Using Multiple Segmentation Methods and 3D Reasoning

In this paper we present a novel algorithm to estimate the surface layout of an indoor scene, which can serve as a visual cue for many different applications, e.g. 3D tracking, or localization in visual odometry. The main contribution of this work lies in combining multiple superpixel segmentation methods in order to obtain semantically meaningful regions. For each segmentation method, we combine 3D reasoning with semantic reasoning to generate multiple surface layout label hypotheses for each pixel. We then get the final label for each pixel within a Markov Random Field (MRF) by combining all hypothesis and by enforcing spatial consistency between neighboring pixels. Experimental results on complex indoor scenes show that our proposed method outperforms state-of-the-art methods.

Exploiting spatial consistency for object classification and pose estimation

In this paper we present a novel object classification and pose recovery algorithm which takes advantage of existing 3D models and multiple synchronized and calibrated views. Having a calibrated scenario provides redundant data which can be exploited for gathering spatial consistency of an objects 3D pose and its class. In a first step, the cameras need to be calibrated and aligned to one common coordinate system. A training set of 3D models, a calibrated setup and Harris corner features are used to find the best fitting 2D projection for an object within the scene. The results are improved by aligning multiple synchronized views to gain spatial consistency. Our experiments using real data show the enhanced results using a calibrated setup over analyzing each camera separately.