Michael Rauter

Reliable Human Detection and Tracking in Top-View Depth Images

The paper presents a method for human detection and tracking in depth images captured by a top-view camera system. We introduce a new feature descriptor which outperforms state-of-the-art features like Simplified Local Ternary Patterns in the given scenario. We use this feature descriptor to train a head-shoulder detector using a discriminative class scheme. A separate processing step ensures that only a minimal but sufficient number of head-shoulder candidates is evaluated. This contributes to an excellent runtime performance. A final tracking step reliably propagates detections in time and provides stable tracking results. The quality of the presented method allows us to recognize many challenging situations with humans tailgating and piggybacking.

GPU-based non-parametric background subtraction for a practical surveillance system

In this paper we present a background subtraction algorithm for a practical surveillance system, on a GPU. It utilizes a compressed non-parametric representation of the history of each pixel, using YCbCr color space, not requiring an offline training period. Although it can be parametrized to cope successfully with moving background, we rather focus on fulfilling some requirements of a practical surveillance system monitoring pedestrians and traffic. First, the time it takes for a stopped foreground object to be absorbed into the background (integration time) should be large enough. Furthermore, the integration time should be controllable by the user and should remain constant, regardless of the complexity of the scene. A further requirement is that objects which repeatedly re-appear in the image, e.g. vehicles having similar colors crossing repeatedly the same region in the image, need not be incorporated into the background. In addition, foreground aperture is undesired, even in case of slowly moving large objects. We implement our method on a NVidia GeForce 9800 GT GPU, achieving 635 fps for the background algorithm, or 436 fps when memory transfer to and from the GPU is included, on a video with 352×288 resolution. We demonstrate the capability of the algorithm by comparing it to MoG, both on moving background and on practical surveillance scenarios. Our method outperforms MoG in both modes, in terms of adaptation speed, run-time and the quality of the foreground segmentation. Furthermore, the integration time is more stable.

Pedestrian detection using GPU-accelerated multiple cue computation

Achieving accurate pedestrian detection for practically relevant scenarios in real-time is an important problem for many applications, while representing a major scientific challenge at the same time. In this paper we present an algorithmic framework which efficiently computes pedestrian-specific shape and motion cues and combines them in a probabilistic manner to infer the location and occlusion status of pedestrians viewed by a stationary camera. The articulated pedestrian shape is represented by a set of sparse contour templates, where fast template matching against image features is carried out using integral images built along oriented scan-lines. The motion cue is obtained by employing a non-parametric background model using the YCbCr color space. Both cues are computed and evaluated on the GPU. Given the probabilistic output from the two cues the spatial configuration of hypothesized human body locations is obtained by an iterative optimization scheme taking into account the depth ordering and occlusion status of individual hypotheses. The method achieves fast computation times even in complex scenarios with a high pedestrian density. Employed computational schemes are described in detail and the validity of the approach is demonstrated on three PETS2009 datasets depicting increasing pedestrian density. Evaluation results and comparison with state of the art are presented.

A GPU accelerated Fast Directional Chamfer Matching algorithm and a detailed comparison with a highly optimized CPU implementation

In this work we present an efficient GPU implementation of the Fast Directional Chamfer Matching (FDCM) algorithm [10]. We propose some extensions to the original FDCM algorithm. In particular, we extend the algorithm to handle templates with variable size, to account for perspective effects. To the best of our knowledge, our work is the first to present a full implementation of a shape based matching algorithm on a GPU. Further contributions of our work consist of implementing a highly optimized CPU version of the algorithm (via multi-threading and SSE2), as well as a thorough comparison between pure GPU, pure CPU, and a hybrid version. The hybrid CPU-GPU version which turns out to be the fastest, achieves run-time of 44 fps on PAL resolution images.

A multisensor surveillance system for Automated Border Control (eGate)

This paper presents1 a multisensor surveillance system used inside an Automated Border Control (ABC) system (more specifically, an eGate). The system consists of two parts: counting the number of persons inside the eGate (person separation), ensuring that no more than one passenger is present; left luggage detection, ensuring that the passenger did not leave any item inside the eGate. These tasks are performed using a top-view mounted sensor inside the eGate, consisting of a trinocular camera configuration comprised of a monochrome stereo setup which delivers depth information, and a color camera mounted in-between, capturing color information. In contrast to already existing ABC solutions, which mostly use electronic sensors, e.g. simple beam technology, for person separation and left item detection, we introduce vision based technologies to elevate the security of such systems to a higher level, also increasing the usability for the border guard. The system achieves real time and had been demonstrated and evaluated at the Vienna International Airport.

GPU-Accelerated Human Detection Using Fast Directional Chamfer Matching

We present a GPU-accelerated, real-time and practical, pedestrian detection system, which efficiently computes pedestrian-specific shape and motion cues and combines them in a probabilistic manner to infer the location and occlusion status of pedestrians viewed by a stationary camera. The articulated pedestrian shape is approximated by a mean contour template, where template matching against an incoming image is carried out using line integral based, Fast Directional Chamfer Matching, employing variable scale templates (hybrid CPU-GPU). The motion cue is obtained by employing a compressed non-parametric background model (GPU). Given the probabilistic output from the two cues, the spatial configuration of hypothesized human body locations is obtained by an iterative optimization scheme taking into account the depth ordering and occlusion status of individual hypotheses. The method achieves fast computation times (32 fps) even in complex scenarios with a high pedestrian density. Employed computational schemes are described in detail and the validity of the approach is demonstrated on three PETS2009 datasets depicting increasing pedestrian density.

A CPU-GPU hybrid people counting system for real-world airport scenarios using arbitrary oblique view cameras

This work1 presents a real-time hybrid CPU-GPU implementation of a practical people counting system, developed for real-world airport scenarios and using the existing airport single cameras. The cameras are characterized by low quality images and are installed in arbitrary oblique viewing angles and heights relative to the ground plane. The scenes are characterized by large field of view, large scale variations of people size, high clutter, and in particular severe occlusions. In addition, people tend to remain long at rest while queuing. Furthermore, real-time performance is required and no elaborate camera calibration is feasible. Our system is based on the fusion of two approaches. The first one is holistic, namely a texture based classification. The second approach utilizes the fast directional Chamfer matching algorithm with variable size ellipse templates to detect heads. Using a probabilistic multi-class SVM classifier for both approaches, the output of the 2 classifier is further fused, yielding a unified prediction.

AVSS2011 demo session: Real-time human detection using fast contour template matching for visual surveillance

Summary form only given. Anthropomatics addresses the symbiosis between humans and machines, focusing on a deeper understanding of the cooperation, interaction and coexistence between humans and machines stimulating and strengthen advanced and deep research in response to the challenges of increasingly smart environments and multimodal access to various complex technical systems. At KIT the Focus Anthropomatics and Robotics - APR has been set up by a number of research groups focusing on the research field of Anthropomatics and Robotics with more than 250 researchers. Modelling humans and their capabilities requires a deep understanding of the principle of biomechanics and kinematics, as well as the underlaying neural control principles and the perceptive and actuatoric system. Modelling and understanding of the sensomotoric mechanisms, learning and developement of skills and cognititve capabilities to enable humans to interact with the world is of high importance to design technical systems operating closely and interactively with humans via various modalities like speech, haptics, vision, grasping and locomotion. Typical research fields are related to active vision, interpretation of scenes and human activities, recognition and tracking technologies multimodal & perceptual user interfaces, understanding and translation of speech. Complementary research needed is related to the retrieval & access and summarization of multimedia data sources, translation of spoken text, context aware learning computers, implicit services and many more. The robotics application field ranges from interactive industrial robotics, service robotic companions, humanoids and medical robotics. In all domains the integrating aspects are focusing on algorithms processing real word data as well as open self-organizing architectures which allow autonomy, skill an