Abstract: In various examples, techniques for using hardware feature trackers in autonomous or semi-autonomous systems are described. Systems and methods are disclosed that use a processor(s) to determine flow vectors associated with pixel locations in a first image. The systems also use the processor(s) to determine a location of a feature point in a second image based at least on one or more of the flow vectors and a subpixel location of the feature point in the first image. In some examples, the processor(s) may include an optical flow accelerator (OFA) that includes a hardware unit storing a lookup table that is used to determine the location of the feature point in the second image. In some examples, the processor(s) may include an OFA to determine the flow vectors and a vision processor to determine the location of the feature point in the second image.

Abstract: Apparatuses, systems, and techniques to generate one or more confidence values associated with one or more objects identified by one or more neural networks. In at least one embodiment, one or more confidence values associated with one or more objects identified by one or more neural networks are generated based on, for example, one or more neural network outputs.

Abstract: In various examples, techniques for using hardware feature trackers in autonomous or semi-autonomous systems are described. Systems and methods are disclosed that use a processor(s) to determine flow vectors associated with pixel locations in a first image. The systems also use the processor(s) to determine a location of a feature point in a second image based at least on one or more of the flow vectors and a subpixel location of the feature point in the first image. In some examples, the processor(s) may include an optical flow accelerator (OFA) that includes a hardware unit storing a lookup table that is used to determine the location of the feature point in the second image. In some examples, the processor(s) may include an OFA to determine the flow vectors and a vision processor to determine the location of the feature point in the second image.

Abstract: In various examples, a technique for estimating depth information for stereo images with reduced estimation inaccuracy by performing depth accuracy assessment. The technique includes generating a depth map associated with a first image in a stereo image pair based on at least on stereo features of the first image. The technique also includes generating a confidence map that represents probabilities of depth values in the generated depth map being accurate based at least on the stereo features for the first image. The technique also includes updating one or more portions of the generated depth map based at least on the confidence map.

Abstract: Apparatuses, systems, and techniques to generate one or more confidence values associated with one or more objects identified by one or more neural networks. In at least one embodiment, one or more confidence values associated with one or more objects identified by one or more neural networks are generated based on, for example, one or more neural network outputs.

Abstract: Systems and methods are disclosed that use a geometric approach to detect objects on a road surface. A set of points within a region of interest between a first frame and a second frame are captured and tracked to determine a difference in location between the set of points in two frames. The first frame may be aligned with the second frame and the first pixel values of the first frame may be compared with the second pixel values of the second frame to generate a disparity image including third pixels. Subsets of the third pixels that have an disparity image value about a first threshold may be combined, and the third pixels may be scored and associated with disparity values for each pixel of the one or more subsets of the third pixels. A bounding shape may be generated based on the scoring that corresponds to the object.

Abstract: A geometric approach may be used to detect objects on a road surface. A set of points within a region of interest between a first frame and a second frame are captured and tracked to determine a difference in location between the set of points in two frames. The first frame may be aligned with the second frame and the first pixel values of the first frame may be compared with the second pixel values of the second frame to generate a disparity image including third pixels. One or more subsets of the third pixels that have a value above a first threshold may be combined, and the third pixels may be scored and associated with disparity values for each pixel of the one or more subsets of the third pixels. A bounding shape may be generated based on the scoring.

Abstract: In various examples, techniques for using hardware feature trackers in autonomous or semi-autonomous systems are described. Systems and methods are disclosed that use a processor(s) to determine flow vectors associated with pixel locations in a first image. The systems also use the processor(s) to determine a location of a feature point in a second image based at least on one or more of the flow vectors and a subpixel location of the feature point in the first image. In some examples, the processor(s) may include an optical flow accelerator (OFA) that includes a hardware unit storing a lookup table that is used to determine the location of the feature point in the second image. In some examples, the processor(s) may include an OFA to determine the flow vectors and a vision processor to determine the location of the feature point in the second image.

Abstract: A hybrid matching approach can be used for computer vision that balances accuracy with speed and resource consumption. Stereoscopic image data can be rectified and downsampled, then analyzed using a semi-global matching (SGM) process. The use of downsampled images greatly reduces time and bandwidth requirements, while providing high accuracy disparity results. These disparity results can be provided as external hints to a fast module that can perform a robust matching process in the time needed for applications such as real time navigation. The external hints can be used, along with potentially other hints, to define a search space for use by the fast module, which can result in higher quality disparity results obtained within specified timing constraints and with limited resources. The disparity results can be used to determine distances to various objects, as may be important for vehicle navigation or robotic task performance.

Abstract: Apparatuses, systems, and techniques to generate one or more confidence values associated with one or more objects identified by one or more neural networks. In at least one embodiment, one or more confidence values associated with one or more objects identified by one or more neural networks are generated based on, for example, one or more neural network outputs.

Abstract: Systems and methods are disclosed that use a geometric approach to detect objects on a road surface. A set of points within a region of interest between a first frame and a second frame are captured and tracked to determine a difference in location between the set of points in two frames. The first frame may be aligned with the second frame and the first pixel values of the first frame may be compared with the second pixel values of the second frame to generate a disparity image including third pixels. One or more subsets of the third pixels that have an disparity image value about a first threshold may be combined, and the third pixels may be scored and associated with disparity values for each pixel of the one or more subsets of the third pixels. A bounding shape may be generated based on the scoring that corresponds to the object.

Abstract: Embodiments of the present invention are directed to methods and systems for performing automatic noise reduction in video. According to one aspect of the invention, a video noise-reducing system is provided consisting of a noise estimator, a motion classifier, two stages of filters, each including a spatial and temporal filter, and a combiner. The system adapts to noise level and to scene content to find at each location in the image a balance of noise reduction and detail preservation. Temporal Infinite Impulse Response (IIR) filtering provides a high level of detail-preserving noise reduction where motion allows, while non linear spatial filtering provides edge-preserving noise reduction in areas where the temporal filter would introduce motion artifacts. A spatial-temporal combiner provides smooth transition and balance between the two filtering modes; this block also enables use of external cues to produce a visually pleasing output based on ambient conditions.

Abstract: Embodiments of the present invention are directed to methods and systems for performing automatic noise reduction in video. According to one aspect of the invention, a video noise-reducing system is provided consisting of a noise estimator, a motion classifier, two stages of filters, each including a spatial and temporal filter, and a combiner. The system adapts to noise level and to scene content to find at each location in the image a balance of noise reduction and detail preservation. Temporal Infinite Impulse Response (IIR) filtering provides a high level of detail-preserving noise reduction where motion allows, while non linear spatial filtering provides edge-preserving noise reduction in areas where the temporal filter would introduce motion artifacts. A spatial-temporal combiner provides smooth transition and balance between the two filtering modes; this block also enables use of external cues to produce a visually pleasing output based on ambient conditions.

Abstract: In one aspect, a method for encoding pictures is provided. The method is applied to each picture in a sequence of pictures, and the method comprises the steps of assigning a pre-decoder buffer removal time to the picture; selecting, for the picture, a number of bits, wherein the time-equivalent of the number of bits is no greater than a difference based on the pre-decoder buffer removal time of the picture and an initial arrival time of the picture into a pre-decoder buffer; and compressing the picture to generate the number of bits. The method may further include the step of allocating a first number of bits for compressing the picture and one or more number of bits for compressing one or more future pictures, wherein the future pictures are in the pre-decoder buffer at the pre-decoder buffer removal time of the current picture.

Abstract: A system (and a method) are disclosed for adaptively selecting quantization parameter for each region of input video signal to be encoded within a video processing system. The system includes a frame partition module, an edge feature detector, a macroblock adaptive quantization energy (AQEnergy) evaluator and a macroblock adaptive quantization parameter selector. The frame partition module partitions a frame of an input video signal into smaller blocks of pixel data. The edge feature detector generates an edge direction histogram for each block to be encoded. The macroblock AQEnergy evaluator receives the edge direction histogram of the block, calculates the AQEnergy of the block and generates the adaptive quantization score (AQScore) of the macroblock. The macroblock adaptive quantization parameter selector selects an appropriate macroblock quantization parameter corresponding to the macroblock AQScore by a combination of programmable scaling and threshold logic.

Abstract: In one aspect, a method for encoding pictures is provided. The method is applied to each picture in a sequence of pictures, and the method comprises the steps of assigning a pre-decoder buffer removal time to the picture; selecting, for the picture, a number of bits, wherein the time-equivalent of the number of bits is no greater than a difference based on the pre-decoder buffer removal time of the picture and an initial arrival time of the picture into a pre-decoder buffer; and compressing the picture to generate the number of bits. The method may further include the step of allocating a first number of bits for compressing the picture and one or more number of bits for compressing one or more future pictures, wherein the future pictures are in the pre-decoder buffer at the pre-decoder buffer removal time of the current picture.

Abstract: In one aspect, a method for encoding pictures is provided. The method is applied to each picture in a sequence of pictures, and the method comprises the steps of assigning a pre-decoder buffer removal time to the picture; selecting, for the picture, a number of bits, wherein the time-equivalent of the number of bits is no greater than a difference based on the pre-decoder buffer removal time of the picture and an initial arrival time of the picture into a pre-decoder buffer; and compressing the picture to generate the number of bits. The method may further include the step of allocating a first number of bits for compressing the picture and one or more number of bits for compressing one or more future pictures, wherein the future pictures are in the pre-decoder buffer at the pre-decoder buffer removal time of the current picture.

Abstract: One method for analyzing a bitstream having a plurality of compressed pictures comprises computing an initial arrival time and a final arrival time of a compressed picture, wherein the initial arrival time is equal to an earlier of the final arrival time of the immediately previous compressed picture or equal to a sum of a fixed time plus a sum of removal delays of all of the compressed pictures between the first compressed picture following the buffering period message and the compressed picture, including the compressed picture, and wherein the final arrival time is equal a sum of the initial arrival time and a time calculated based on the number of bits associated with the compressed picture at the bit rate; and verifying that a difference between the final removal time and the initial arrival time does not exceed the time for reaching the buffer size at the bit rate.

Abstract: A video quantizer provides for performing quantization adaptively in accordance with perceptual masking characteristics of the human visual system. In a preferred MPEG encoder-IC, a block-based activity quantization modification or “activity-modification” is formed from the combined correlation of block-energy and edge analyses. A luminance-sensitivity modification is then formed and correlated with the activity modification to form an intermediate modification. A nominal-quantization modification is further formed and correlated with the intermediate modification, which is then limited and correlated with a nominal quantization value to form a base modification. Next, a positional-sensitivity modification is formed as a perimeter offset, which offset is correlated with the base modification to form a modified quantization value, and which modified quantization value is then rounded and returned to a rate controller.

Abstract: In one aspect, a method for encoding pictures is provided. The method is applied to each picture in a sequence of pictures, and the method comprises the steps of assigning a pre-decoder buffer removal time to the picture; selecting, for the picture, a number of bits, wherein the time-equivalent of the number of bits is no greater than a difference based on the pre-decoder buffer removal time of the picture and an initial arrival time of the picture into a pre-decoder buffer; and compressing the picture to generate the number of bits. The method may further include the step of allocating a first number of bits for compressing the picture and one or more number of bits for compressing one or more future pictures, wherein the future pictures are in the pre-decoder buffer at the pre-decoder buffer removal time of the current picture.