Multiple Sensor Fusion for Detection, Classification and Tracking of Moving Objects in Driving Environments
Citations
124 citations
38 citations
Cites background or methods from "Multiple Sensor Fusion for Detectio..."
...The data is then filtered and an appropriate fusion technology implemented this is fed into localization and mapping techniques like SLAM; the same data can be used to identify static or moving objects in the environment and this data can be used to classify the objects, wherein classification information is used to finalize information in creating a model of the environment which in turn can be fed into the control algorithm [27]....
[...]
...Another reason could be the system failure risk due to the failure of that single sensor [21,27,40] and hence one should introduce a level of redundancy....
[...]
26 citations
14 citations
11 citations
Cites background or methods from "Multiple Sensor Fusion for Detectio..."
...Kalman Filter (KF): KF features make it suited to deal with multi-sensor estimation and data fusion problems [11]....
[...]
...Multi-sensor data fusion is the process of combining several observations from different sensor inputs to provide a more complete, robust and precise representation of the environment of interest [11]....
[...]
...Evidence Theory (ET): The advantage of ET is its ability to represent incomplete evidence, total ignorance and the lack of a need for a priori probabilities [11]....
[...]
...Monte Carlo (MC) Methods: MC methods are well suited for problems where state transition models and observation models are highly non-linear [11]....
[...]
...information fusion are [11] complexity (need large number of probabilities), inconsistency (difficult to specify consistent set of beliefs in terms of probability) and model precision (precise probabilities about almost unknown events)....
[...]
References
31,952 citations
"Multiple Sensor Fusion for Detectio..." refers background or methods in this paper
...Dalal and Triggs (2005) present a human classification scheme that uses SIFT-inspired features, called histograms of oriented gradients (HOG), and a linear SVM as a learning method. An HOG feature divides the region into k orientation bins, also defines four different cells that divide the rectangular feature, and then a Gaussian mask is applied to the magnitude values in order to weight the center pixels, and the pixels are interpolated with respect to pixel location within a block. The resulting feature is a vector of dimension 36 containing the summed magnitude of each pixel cells, divided into 9 (kvalue) bins. These features have been extensively exploited in the literature (Dollár et al., 2012). Recently, Qiang et al. (2006) and Chavez-Garcia et al. (2013) use HOG as a weak rule for AdaBoost classifier, achieving the same detection performance, but with less computation time....
[...]
...Dalal and Triggs (2005) present a human classification scheme that uses SIFT-inspired features, called histograms of oriented gradients (HOG), and a linear SVM as a learning method....
[...]
...Each element of a vector is a histogram of gradient orientations (Dalal and Triggs, 2005)....
[...]
...Dalal and Triggs (2005) present a human classification scheme that uses SIFT-inspired features, called histograms of oriented gradients (HOG), and a linear SVM as a learning method. An HOG feature divides the region into k orientation bins, also defines four different cells that divide the rectangular feature, and then a Gaussian mask is applied to the magnitude values in order to weight the center pixels, and the pixels are interpolated with respect to pixel location within a block. The resulting feature is a vector of dimension 36 containing the summed magnitude of each pixel cells, divided into 9 (kvalue) bins. These features have been extensively exploited in the literature (Dollár et al., 2012). Recently, Qiang et al. (2006) and Chavez-Garcia et al. (2013) use HOG as a weak rule for AdaBoost classifier, achieving the same detection performance, but with less computation time. Maji et al. (2008) and Wu and Nevatia (2007) proposed a feature based on segments of lines or curves, and compared it with HOG using AdaBoost and SVM learning algorithms....
[...]
...We based our visual representation approach on the work of Dalal and Triggs (2005) on histograms of oriented gradients (HOG) which has recently become a stateof-the-art feature in the computer vision domain for object detection tasks....
[...]
18,620 citations
15,391 citations
"Multiple Sensor Fusion for Detectio..." refers background or methods in this paper
...Moreover, we do not need a data association process to relate the moving objects from lidar and camera because the camera uses ROI from lidar processing. Afterwards, the fused mass distribution is considered as the reference distribution and therefore combined with the radar mass assignment mr. The association between lidar and radar objects is done using a gating approach between tracks based on the covariance matrices of the tracks from both sensors, this approach is based on the association techniques proposed by Bar-Shalom and Tse (1975) and Baig (2012). Also we include the idea of associating tracks that have parallel trajectories to perform track confirmation....
[...]
...Usually, the fixed size of a pedestrian is modified during ROI generation using size factors. This idea has two main drawbacks: the number of candidates can be very large, which makes it difficult to fulfil real-time requirements; and many irrelevant regions are passed to the next module, which increases the potential number of false positives. As a result, other approaches are used to perform explicit segmentation based on camera image, road restriction, or complementary sensor measurements. The most robust techniques to generate ROIs using camera data are biologically inspired. Milch and Behrens (2001) select ROIs according to color, intensity and gradient orientation of pixels....
[...]
...Regarding stereo-based pose estimation, Labayrade et al. (2007) introduced v-disparity space, which consists of accumulating stereo disparity along the image y-axis in order to compute the slope of the road and to point out the existence of vertical objects when the accumulated disparity of an image row is very different from its neighbors....
[...]
...Usually, the fixed size of a pedestrian is modified during ROI generation using size factors. This idea has two main drawbacks: the number of candidates can be very large, which makes it difficult to fulfil real-time requirements; and many irrelevant regions are passed to the next module, which increases the potential number of false positives. As a result, other approaches are used to perform explicit segmentation based on camera image, road restriction, or complementary sensor measurements. The most robust techniques to generate ROIs using camera data are biologically inspired. Milch and Behrens (2001) select ROIs according to color, intensity and gradient orientation of pixels. Dollár et al. (2012) review in detail current state-of-the-art intensity-based hypothesis generation for pedestrian detection. Some of these methods involve the use of learning techniques to discover threshold values for intensity segmentation. Optical flow has been used for foreground segmentation, specially in the general context of moving obstacle detection. Franke, U. and Heinrich (2002) propose to merge stereo processing, which extracts depth information without time correlation, and motion analysis, which is able to detect small gray value changes in order to permit early detection of moving objects, e....
[...]
...Regarding stereo-based pose estimation, Labayrade et al. (2007) introduced v-disparity space, which consists of accumulating stereo disparity along the image y-axis in order to compute the slope of the road and to point out the existence of vertical objects when the accumulated disparity of an image row is very different from its neighbors. Andriluka et al. (2008) proposed fitting 3D road data points to a plane, whereas Singh et al....
[...]
14,565 citations
13,037 citations