Abstract: An example apparatus for selecting priors includes a training set receiver to receive a training dataset. The apparatus includes a prior generator to generate a set of redundant priors based on the training dataset. The apparatus includes an intermediate trainer to train a detection CNN using the set of redundant priors. The apparatus includes a score and location receiver to send all training samples of the training dataset to the trained detection CNN and receive responses for all of the redundant priors in the set of redundant priors. The apparatus includes a subset selector to select a subset of the set of redundant priors based on the responses.

Abstract: An example apparatus for selecting priors includes a training set receiver to receive a training dataset. The apparatus includes a prior generator to generate a set of redundant priors based on the training dataset. The apparatus includes an intermediate trainer to train a detection CNN using the set of redundant priors. The apparatus includes a score and location receiver to send all training samples of the training dataset to the trained detection CNN and receive responses for all of the redundant priors in the set of redundant priors. The apparatus includes a subset selector to select a subset of the set of redundant priors based on the responses.

Abstract: A tracking algorithm includes a tracking failure detection technique using a key points selection algorithm. As a result, the described system is able to work on an embedded device in real time, providing high quality vehicle detection and tracking, in some embodiments. A vision system detects and tracks vehicles from sequences of images taken from another moving vehicle in real-time on an embedded platform. The system can achieve real-time performance on an embedded platform in some embodiments, taking into account that modern boosting detectors are too slow for use in such a system. A tracker includes an algorithm for tracking and an algorithm for detection of tracking failures. The tracking algorithm is based on an optical flow calculation for key points selected based on the distribution of features from the last detection. The algorithm for detection of tracking failures is based on an estimation of low confidence detections.

Abstract: A two-stage pipeline includes a first stage that is a static filter applied to an inverse perspective mapping (IPM) image. A SLAT (Symmetrical Local Adaptive Threshold) method, according to some embodiments, is more adaptive to environment changes. Additional geometry restrictions in the static filter remove clutter. The second stage of the pipeline is a dynamic filter. The motion of blobs is analyzed to find regions that look and behave like road marking.

Abstract: In accordance with some embodiments Adaptive Channel Features may be implemented by determining random features. The random features may be determined by defining a maximum allowed feature size of training samples. Then random filter positions of a training sample are sampled. Thereafter, pixel weights in a patch of the maximum allowed feature size is calculated. A feature is selected for applying a boosted classifier.

Abstract: A camera facing the front of a vehicle while the vehicle is moving on the road may be calibrated by receiving sequential images from the camera. Image key points in the area limited by the road location are selected. The key points are tracked using an optical flow method. A filtering procedure is applied to the key points to identify the straight-line motion of the vehicle. At least two straight lines corresponding to opposite sides of the road. A calibration algorithm is applied to the at least two lines to determine a vanishing point. The pitch and/or yaw angles of the camera are then calculated.

Abstract: A specified object in still images or video may be detected using a sliding search window technique, applied to the original image and its downscaled versions, in order to detect objects of different sizes. At each scale and each position of the sliding window, the technique may use a boosted tree classifier to determine whether the window contains the object. It may exit earlier if some intermediate sum falls below the certain threshold. To accelerate object detection, hybrid features are used in different boosted chains. For first boosted chains, the fastest features may be applied and then, after more complex (but slower) features and for the last few chains, the most powerful feature (but most computationally expensive) is used. This strategy may improve the speed of detection because for a majority of checking windows, only first boosted chains are used and so only the fastest features are calculated in some embodiments.

Abstract: A camera facing the front of a vehicle while the vehicle is moving on the road may be calibrated by receiving sequential images from the camera. Image key points in the area limited by the road location are selected. The key points are tracked using an optical flow method. A filtering procedure is applied to the key points to identify the straight-line motion of the vehicle. At least two straight lines corresponding to opposite sides of the road. A calibration algorithm is applied to the at least two lines to determine a vanishing point. The pitch and/or yaw angles of the camera are then calculated.

Abstract: In accordance with some embodiments Adaptive Channel Features may be implemented by determining random features. The random features may be determined by defining a maximum allowed feature size of training samples. Then random filter positions of a training sample are sampled. Thereafter, pixel weights in a patch of the maximum allowed feature size is calculated. A feature is selected for applying a boosted classifier.

Abstract: A tracking algorithm includes a tracking failure detection technique using a key points selection algorithm. As a result, the described system is able to work on an embedded device in real time, providing high quality vehicle detection and tracking, in some embodiments. A vision system detects and tracks vehicles from sequences of images taken from another moving vehicle in real-time on an embedded platform. The system can achieve real-time performance on an embedded platform in some embodiments, taking into account that modern boosting detectors are too slow for use in such a system. A tracker includes an algorithm for tracking and an algorithm for detection of tracking failures. The tracking algorithm is based on an optical flow calculation for key points selected based on the distribution of features from the last detection. The algorithm for detection of tracking failures is based on an estimation of low confidence detections.

Abstract: A two-stage pipeline includes a first stage that is a static filter applied to an inverse perspective mapping (IPM) image. A SLAT (Symmetrical Local Adaptive Threshold) method, according to some embodiments, is more adaptive to environment changes. Additional geometry restrictions in the static filter remove clutter. The second stage of the pipeline is a dynamic filter. The motion of blobs is analyzed to find regions that look and behave like road marking.

Abstract: A specified object in still images or video may be detected using a sliding search window technique, applied to the original image and its downscaled versions, in order to detect objects of different sizes. At each scale and each position of the sliding window, the technique may use a boosted tree classifier to determine whether the window contains the object. It may exit earlier if some intermediate sum falls below the certain threshold. To accelerate object detection, hybrid features are used in different boosted chains. For first boosted chains, the fastest features may be applied and then, after more complex (but slower) features and for the last few chains, the most powerful feature (but most computationally expensive) is used. This strategy may improve the speed of detection because for a majority of checking windows, only first boosted chains are used and so only the fastest features are calculated in some embodiments.

Abstract: Estimation of a 3D layout of roads and paths traveled by pedestrians is achieved by observing the pedestrians and estimating road parameters from the pedestrian's size and position in a sequence of video frames. The system includes a foreground object detection unit to analyze video frames of a 3D scene and detect objects and object positions in video frames, an object scale prediction unit to estimate 3D transformation parameters for the objects and to predict heights of the objects based at least in part on the parameters, and a road map detection unit to estimate road boundaries of the 3D scene using the object positions to generate the road map.

Abstract: Estimation of a 3D layout of roads and paths traveled by pedestrians is achieved by observing the pedestrians and estimating road parameters from the pedestrian's size and position in a sequence of video frames. The system includes a foreground object detection unit to analyze video frames of a 3D scene and detect objects and object positions in video frames, an object scale prediction unit to estimate 3D transformation parameters for the objects and to predict heights of the objects based at least in part on the parameters, and a road map detection unit to estimate road boundaries of the 3D scene using the object positions to generate the road map.

Abstract: In an embodiment, an image is received having a first portion and one or more other portions. The one or more other portions are replaced with one or more other images. The replacing of the one or more portions results in an image including the first portion and the one or more other images. In an embodiment, the background of an image is replaced with another background. In an embodiment, the foreground is extracted by identifying the background based on an image of the background without any foreground. In an embodiment, the foreground is extracted by identifying portions of the image that have characteristics that are expected to be associated with the background and characteristics that are expected to be associated with foreground. In an embodiment any of the images can be still images. In an embodiment, any of the images are video images.