Citation Context

...s from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields [10, 28] or gradient histograms (HOG) =-=[2, 3, 26, 29, 31]-=-. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or AdaBoost cascades [16, 24, 25, 30, 31]. Besides ...

Citation Context

...erms of pedestrian shape, size and location is utilized. Third, we estimate the degree of visibility of each component given the selected cluster. For segmentation, we chose the mean-shift algorithm, =-=[1]-=-, out of many possible choices. As shown in [8], meanshift provides a good balance between segmentation accuracy and processing efficiency. The result of the mean-shift segmentation is a set of C clus...

Citation Context

...y P (ω0|xi) is approximated using a component-based mixture-of-experts model. A sample xi is composed out of K components which are usually related to body parts. In the mixture-of-experts framework, =-=[13]-=-, the final decision results from a weighted linear combination of so-called local expert classifiers which are specialized in a particular area of the feature space. With Fk(xi) representing a local ...

Citation Context

...aches for pedestrian classification. See [6] for a current survey. Recently there have been efforts to break down the complexity of pedestrian appearance into components usually related to body parts =-=[5, 9, 14, 15, 16, 18, 22, 23, 26, 30]-=-. After detecting the individual body parts, detection results are fused using statistical models [15, 16, 30], learning or voting schemes [5, 14, 18, 22] or heuristics [26]. In view of detecting part...

Citation Context

...pic recently. Most state-of-the art systems, cf. [4, 6, 12], derive a set of features from the available image data and apply pattern classification techniques. Popular features include Haar wavelets =-=[18, 20, 25]-=-, adaptive local receptive fields [10, 28] or gradient histograms (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector ...

Citation Context

...ors have proposed multi-cue approaches combining information from different modalities, e.g. intensity, depth and motion [7, 10, 29]. We do not consider work in the domain of 3D human pose estimation =-=[17]-=-, but focus on discriminative 2D approaches for pedestrian classification. See [6] for a current survey. Recently there have been efforts to break down the complexity of pedestrian appearance into com...

Citation Context

...pic recently. Most state-of-the art systems, cf. [4, 6, 12], derive a set of features from the available image data and apply pattern classification techniques. Popular features include Haar wavelets =-=[18, 20, 25]-=-, adaptive local receptive fields [10, 28] or gradient histograms (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector ...

Citation Context

...pic recently. Most state-of-the art systems, cf. [4, 6, 12], derive a set of features from the available image data and apply pattern classification techniques. Popular features include Haar wavelets =-=[18, 20, 25]-=-, adaptive local receptive fields [10, 28] or gradient histograms (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector ...

Citation Context

...grams (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or AdaBoost cascades =-=[16, 24, 25, 30, 31]-=-. Besides operating in the image intensity domain only, some authors have proposed multi-cue approaches combining information from different modalities, e.g. intensity, depth and motion [7, 10, 29]. W...

Citation Context

...aches for pedestrian classification. See [6] for a current survey. Recently there have been efforts to break down the complexity of pedestrian appearance into components usually related to body parts =-=[5, 9, 14, 15, 16, 18, 22, 23, 26, 30]-=-. After detecting the individual body parts, detection results are fused using statistical models [15, 16, 30], learning or voting schemes [5, 14, 18, 22] or heuristics [26]. In view of detecting part...

Citation Context

...s from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields [10, 28] or gradient histograms (HOG) =-=[2, 3, 26, 29, 31]-=-. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or AdaBoost cascades [16, 24, 25, 30, 31]. Besides ...

Citation Context

...n-pedestrian (ω1) samples xi ∈D. Each sample xi =[xi i ; xdi ; xfi ] consists of three different modalities, i.e. gray-level image intensity (xi i ), dense depth information via stereo vision (xd i ) =-=[11]-=- and dense optical flow (xfi ) [27]. We treat xd i and xfi similarly to gray-level intensity images xii , in that both depth and motion cues are represented as images, where pixel values encode distan...

Citation Context

...s from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields [10, 28] or gradient histograms (HOG) =-=[2, 3, 26, 29, 31]-=-. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or AdaBoost cascades [16, 24, 25, 30, 31]. Besides ...

Citation Context

...grams (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or AdaBoost cascades =-=[16, 24, 25, 30, 31]-=-. Besides operating in the image intensity domain only, some authors have proposed multi-cue approaches combining information from different modalities, e.g. intensity, depth and motion [7, 10, 29]. W...

Citation Context

...estrian classification. Since we require partially occluded multicue (intensity, dense stereo, dense optical flow) training and test samples, we cannot use established datasets for benchmarking, e.g. =-=[2, 4, 6, 19]-=-. To our knowledge, our dataset is the first to comprise “real” partially occluded pedestrians in the field of pedestrian classification. Wang et al. only simulated partial occlusion by synthetically ...

Citation Context

...ians. If some body parts of a pedestrian are occluded, the classification results often do not degrade gracefully. Component-based approaches which represent a pedestrian as an ensemble of parts, cf. =-=[6]-=-, can only alleviate this problem to some extent without prior knowledge. The key to successful detection of partially occluded pedestrians is additional information about which body parts are occlude...

Citation Context

...fits obtained from the proposed mixture-of-experts framework. 2. Previous Work Pedestrian classification has become an increasingly popular research topic recently. Most state-of-the art systems, cf. =-=[4, 6, 12]-=-, derive a set of features from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields [10, 28] or g...

Citation Context

...grams (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or AdaBoost cascades =-=[16, 24, 25, 30, 31]-=-. Besides operating in the image intensity domain only, some authors have proposed multi-cue approaches combining information from different modalities, e.g. intensity, depth and motion [7, 10, 29]. W...

Citation Context

...f. [4, 6, 12], derive a set of features from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields =-=[10, 28]-=- or gradient histograms (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or ...

Citation Context

...search purposes. 1 We chose to evaluate our approach in a pedestrian classification setting, where we assume that initial pedestrian location hypotheses already exist, e.g. using methods described in =-=[6, 7, 10, 22]-=- or non-vision sensors. In our experiments, we focus on the central part of a pedestrian detection system, the classifier, to eliminate auxiliary effects arising from various detector parameters such ...

Citation Context

...aches for pedestrian classification. See [6] for a current survey. Recently there have been efforts to break down the complexity of pedestrian appearance into components usually related to body parts =-=[5, 9, 14, 15, 16, 18, 22, 23, 26, 30]-=-. After detecting the individual body parts, detection results are fused using statistical models [15, 16, 30], learning or voting schemes [5, 14, 18, 22] or heuristics [26]. In view of detecting part...

Citation Context

...s from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields [10, 28] or gradient histograms (HOG) =-=[2, 3, 26, 29, 31]-=-. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or AdaBoost cascades [16, 24, 25, 30, 31]. Besides ...

Citation Context

...f. [4, 6, 12], derive a set of features from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields =-=[10, 28]-=- or gradient histograms (HOG) [2, 3, 26, 29, 31]. These features are combined with a variety of classifiers, such as neural networks [10, 28], support vector machines (SVMs) [2, 3, 18, 20, 26, 31] or ...

Citation Context

...ch sample xi =[xi i ; xdi ; xfi ] consists of three different modalities, i.e. gray-level image intensity (xi i ), dense depth information via stereo vision (xd i ) [11] and dense optical flow (xfi ) =-=[27]-=-. We treat xd i and xfi similarly to gray-level intensity images xii , in that both depth and motion cues are represented as images, where pixel values encode distance from the camera and magnitude of...

Citation Context

...aches for pedestrian classification. See [6] for a current survey. Recently there have been efforts to break down the complexity of pedestrian appearance into components usually related to body parts =-=[5, 9, 14, 15, 16, 18, 22, 23, 26, 30]-=-. After detecting the individual body parts, detection results are fused using statistical models [15, 16, 30], learning or voting schemes [5, 14, 18, 22] or heuristics [26]. In view of detecting part...

Citation Context

...aches for pedestrian classification. See [6] for a current survey. Recently there have been efforts to break down the complexity of pedestrian appearance into components usually related to body parts =-=[5, 9, 14, 15, 16, 18, 22, 23, 26, 30]-=-. After detecting the individual body parts, detection results are fused using statistical models [15, 16, 30], learning or voting schemes [5, 14, 18, 22] or heuristics [26]. In view of detecting part...

Citation Context

...estrian standing upright on the ground) and move uniformly. We employ a three-step procedure to derive component weights wk(xi) from an unseen sample xi: First, we apply a segmentation algorithm, cf. =-=[8]-=-, to the dense stereo(a) (b) Figure 3. (a) Probability masks for front/back, left and right view. The values of the probability masks are in the range of zero (dark blue) to one (dark red). The value...

Citation Context

...pproach. Off-line, we train multi-cue component-based expert classifiers involving feature spaces derived from gray-level images, depth maps (dense stereo vision) and motion (dense optical flow), cf. =-=[3, 7, 21, 29]-=-. On-line, we apply multi-cue (depth and motion) meanshift segmentation to each test sample to recover occlusiondependent component weights which are used to fuse the component-based expert classifier...

Citation Context

...fits obtained from the proposed mixture-of-experts framework. 2. Previous Work Pedestrian classification has become an increasingly popular research topic recently. Most state-of-the art systems, cf. =-=[4, 6, 12]-=-, derive a set of features from the available image data and apply pattern classification techniques. Popular features include Haar wavelets [18, 20, 25], adaptive local receptive fields [10, 28] or g...

Citation Context

...aches for pedestrian classification. See [6] for a current survey. Recently there have been efforts to break down the complexity of pedestrian appearance into components usually related to body parts =-=[5, 9, 14, 15, 16, 18, 22, 23, 26, 30]-=-. After detecting the individual body parts, detection results are fused using statistical models [15, 16, 30], learning or voting schemes [5, 14, 18, 22] or heuristics [26]. In view of detecting part...