Abstract: Techniques related to training and implementing convolutional neural networks for object recognition are discussed. Such techniques may include applying, at a first convolutional layer of the convolutional neural network, 3D filters of different spatial sizes to an 3D input image segment to generate multi-scale feature maps such that each feature map has a pathway to fully connected layers of the convolutional neural network, which generate object recognition data corresponding to the 3D input image segment.

Abstract: Techniques related to training and implementing convolutional neural networks for object recognition are discussed. Such techniques may include applying, at a first convolutional layer of the convolutional neural network, 3D filters of different spatial sizes to an 3D input image segment to generate multi-scale feature maps such that each feature map has a pathway to fully connected layers of the convolutional neural network, which generate object recognition data corresponding to the 3D input image segment.

Abstract: Methods and apparatus relating to FPGA (Field-Programmable Gate Array) based acceleration in robot motion planning are described. In an embodiment, logic circuitry (such as an FPGA), coupled to a processor, accelerates one or more motion planning operations for a plurality of objects. A first memory, coupled to the logic circuitry, stores data corresponding to a plurality of Oriented Bounding Boxes (OBBs). The plurality of OBBs are to provide Bounding Volume (BV) models for the plurality of objects. Other embodiments are also disclosed and claimed.

Abstract: A technology is described for mapping a physical environment. An example method may include receiving laser point data for laser light reflected from the physical environment and detected by a laser sensor. Points included in the laser point data can be correlated to grid cells in an environment map that represents the physical environment. Error ranges for the points correlated to the grid cells can be determined based in part on an error distribution. Occupation probabilities can then be calculated for the grid cells in the environment map using an interpolation technique and grid cell occupation probabilities for adjacent error grid cells selected based in part on the error ranges of the points, and the grid cells in the environment map can be updated with the occupation probabilities.

Abstract: Systems, apparatuses and methods may provide for technology that generates a sequence of predictive maps based on a sequence of historical maps and overlays the sequence of predictive maps on one another to obtain a map overlay. The technology may also apply an attenuation factor to the map overlay. In one example, the map overlay includes a grid of cells and each cell includes an occupation probability in accordance with the attenuation factor.

Abstract: Methods, apparatus, systems and articles of manufacture are disclosed to improve resource utilization for binary tree structures. An example apparatus to improve resource utilization for field programmable gate array (FPGA) resources includes a computation determiner to identify a computation capability value associated with the FPGA resources, a k-ary tree builder to build a first k-ary tree having a number of k-ary nodes equal to the computation capability value, and an FPGA memory controller to initiate collision computation by transferring the first k-ary tree to a first memory of the FPGA resources.

Abstract: Methods and apparatus relating to FPGA (Field-Programmable Gate Array) based acceleration in robot motion planning are described. In an embodiment, logic circuitry (such as an FPGA), coupled to a processor, accelerates one or more motion planning operations for a plurality of objects. A first memory, coupled to the logic circuitry, stores data corresponding to a plurality of Oriented Bounding Boxes (OBBs). The plurality of OBBs are to provide Bounding Volume (BV) models for the plurality of objects. Other embodiments are also disclosed and claimed.

Abstract: Techniques related to training and implementing convolutional neural networks for object recognition are discussed. Such techniques may include applying, at a first convolutional layer of the convolutional neural network, 3D filters of different spatial sizes to an 3D input image segment to generate multi-scale feature maps such that each feature map has a pathway to fully connected layers of the convolutional neural network, which generate object recognition data corresponding to the 3D input image segment.

Abstract: Methods, apparatus, systems and articles of manufacture are disclosed to improve resource utilization for binary tree structures. An example apparatus to improve resource utilization for field programmable gate array (FPGA) resources includes a computation determiner to identify a computation capability value associated with the FPGA resources, a k-ary tree builder to build a first k-ary tree having a number of k-ary nodes equal to the computation capability value, and an FPGA memory controller to initiate collision computation by transferring the first k-ary tree to a first memory of the FPGA resources.

Abstract: Systems, apparatuses and methods may provide for technology that generates a sequence of predictive maps based on a sequence of historical maps and overlays the sequence of predictive maps on one another to obtain a map overlay. The technology may also apply an attenuation factor to the map overlay. In one example, the map overlay includes a grid of cells and each cell includes an occupation probability in accordance with the attenuation factor.

Abstract: A technology is described for mapping a physical environment. An example method may include receiving laser point data for laser light reflected from the physical environment and detected by a laser sensor. Points included in the laser point data can be correlated to grid cells in an environment map that represents the physical environment. Error ranges for the points correlated to the grid cells can be determined based in part on an error distribution. Occupation probabilities can then be calculated for the grid cells in the environment map using an interpolation technique and grid cell occupation probabilities for adjacent error grid cells selected based in part on the error ranges of the points, and the grid cells in the environment map can be updated with the occupation probabilities.

Abstract: Disclosed is an object detection method used to detect an object in an image pair corresponding to a current frame. The image pair includes an original image of the current frame and a disparity map of the same current frame. The original image of the current frame includes at least one of a grayscale image and a color image of the current frame. The object detection method comprises steps of obtaining a first detection object detected in the disparity map of the current frame; acquiring an original detection object detected in the original image of the current frame; correcting, based on the original detection object, the first detection object so as to obtain a second detection object; and outputting the second detection object.

Abstract: Disclosed are a road line detection method and a road line detection device. The road line detection method comprises a step of obtaining a first disparity map including one or more road regions and a corresponding V-disparity image; a step of sequentially detecting plural sloped line segments in the corresponding V-disparity image according to a big-to-small order of disparities and a big-to-small order of V-values, to serve as plural sequentially adjacent road surfaces; a step of obtaining a second disparity map of plural road line regions of interest corresponding to the plural sloped line segments; and a step of detecting one or more road lines in the second disparity map of the plural road line regions of interest.

Abstract: A method and an apparatus are used for detecting a road partition, the method comprising: a step of obtaining a disparity top view having a road area; and a step of detecting parallel lines as the road partitions from the disparity top view.

Abstract: Disclosed is an object tracking method. The method includes steps of obtaining a first boundary region of a waiting-for-recognition object in the disparity map related to the current frame; calculating a probability of each valid pixel in the first boundary region so as to get a pixel probability map of the waiting-for-recognition object; obtaining historic tracking data of each tracked object, which includes identifier information of the tracked object and a pixel probability map related to each of one or more prior frame related disparity maps prior to the disparity map related to the current frame; determining identifier information of the waiting-for-recognition object, and updating the pixel probability map of the waiting-for-recognition object; and updating, based on the updated pixel probability map of the waiting-for-recognition object, the first boundary region of the waiting-for-recognition object, so as to get a second boundary region of the waiting-for-recognition object.

Abstract: A method and a device are disclosed for detecting a drivable region of a road, the method comprising the steps of: deriving a disparity map from a gray-scale map including the road and detecting the road from the disparity map; removing a part with a height above the road greater than a predetermined height threshold from the disparity map so as to generate a sub-disparity map; converting the sub-disparity map into a U-disparity map; detecting the drivable region from the U-disparity map; and converting the drivable region detected from the U-disparity map into the drivable region within the gray-scale map.

Abstract: Disclosed is an object detection method used to detect an object in an image pair corresponding to a current frame. The image pair includes an original image of the current frame and a disparity map of the same current frame. The original image of the current frame includes at least one of a grayscale image and a color image of the current frame. The object detection method comprises steps of obtaining a first detection object detected in the disparity map of the current frame; acquiring an original detection object detected in the original image of the current frame; correcting, based on the original detection object, the first detection object so as to obtain a second detection object; and outputting the second detection object.

Abstract: Disclosed is an object tracking method. The method includes steps of obtaining a first boundary region of a waiting-for-recognition object in the disparity map related to the current frame; calculating a probability of each valid pixel in the first boundary region so as to get a pixel probability map of the waiting-for-recognition object; obtaining historic tracking data of each tracked object, which includes identifier information of the tracked object and a pixel probability map related to each of one or more prior frame related disparity maps prior to the disparity map related to the current frame; determining identifier information of the waiting-for-recognition object, and updating the pixel probability map of the waiting-for-recognition object; and updating, based on the updated pixel probability map of the waiting-for-recognition object, the first boundary region of the waiting-for-recognition object, so as to get a second boundary region of the waiting-for-recognition object.

Abstract: Disclosed are a method and a system for detecting an object on a road. The method comprises a step of simultaneously capturing two depth maps, and then calculating a disparity map; a step of obtaining, based on the disparity map, a V-disparity image by adopting a V-disparity algorithm; a step of detecting an oblique line in the V-disparity image, and then removing points in the disparity map, corresponding to the oblique line so as to acquire a sub-disparity map excluding the road; a step of detecting plural vertical lines in the V-disparity image, and then extracting, for each of the plural vertical lines, points corresponding to this vertical line from the sub-disparity map as an object sub-disparity map corresponding to this vertical line; and a step of merging any two rectangular areas of the object sub-disparity maps approaching each other, into a rectangular object area.

Abstract: Disclosed are a road line detection method and a road line detection device. The road line detection method comprises a step of obtaining a first disparity map including one or more road regions and a corresponding V-disparity image; a step of sequentially detecting plural sloped line segments in the corresponding V-disparity image according to a big-to-small order of disparities and a big-to-small order of V-values, to serve as plural sequentially adjacent road surfaces; a step of obtaining a second disparity map of plural road line regions of interest corresponding to the plural sloped line segments; and a step of detecting one or more road lines in the second disparity map of the plural road line regions of interest.