A vehicle detection method includes (1) vehicle likelihood region identifying step; (2) vehicle component locating step; and (3) vehicle detecting step. To reduce complexity of calculation and enhance accuracy of detection, the method uses a vehicle likelihood region identifying algorithm to eliminate background regions from a total thermal image and keep vehicle likelihood regions therein for use in further analysis and processing, detects obvious vehicle components, such as vehicle windows and vehicle bottoms, in the thermal image to thereby identify vehicle component likelihood regions, describes a space geometric relationship of vehicle components with a Markov random field model, defines vehicle detection as problems with maximum a posteriori probability, estimates the most likely configuration with an optimization algorithm, so as to effectuate vehicle detection.

A method of tracking an object of an object tracking apparatus is provided. The method includes by performing probability propagation between a set of frames where tracking of a target object is completed and a set of frames where tracking of a target object is not completed among a plurality of frames of an image, calculating a probability map for a target object in each frame included the set of frames where tracking is completed, selecting a frame form the set of frames where tracking is not completed based on the calculated probability map, and determining a location of the target object in the selected frame.

Computer vision systems and methods for end-to end training of neural networks are provided. The system generates a fixed point algorithm for dual-decomposition of a maximum-a-posteriori inference problem and trains the convolutional neural network and a conditional random field with the fixed point algorithm and a plurality of images of a dataset to learn to perform semantic image segmentation. The system can segment an attribute of an image of the dataset by the trained neural network and the conditional random field.

Different from prior works that model the internal distribution of patches within an image implicitly with a top-down latent variable model (e.g., generator), embodiments explicitly represent the statistical distribution within a single image by using an energy-based generative framework, where a pyramid of energy functions, each parameterized by a bottom-up deep neural network, are used to capture the distributions of patches at different resolutions. Also, embodiments of a coarse-to-fine sequential training and sampling strategy are presented to train the model efficiently. Besides learning to generate random samples from white noise, embodiments can learn in parallel with a self-supervised task (e.g., recover an input image from its corrupted version), which can further improve the descriptive power of the learned model. Embodiments does not require an auxiliary model (e.g., discriminator) to assist the training, and embodiments also unify internal statistics learning and image generation in a single framework.

The invention relates to a method for analyzing a movement by a human being, the method including the following steps: selecting an initial probability function, processing, recognizing the end of the movement to be analyzed when a criterion is verified, the processing step being iterated for as long as the criterion is not verified, determining a piece of information regarding the movement to be analyzed, each iteration of the processing step comprising a step for: providing a set of characteristic parameters relative to the movement to be analyzed during the chosen computing time interval, computing a point, the computed point belonging to a reference kinematic and making a function depending on the a posteriori conditional probability function extremal, the determination step taking into account the points computed upon each iteration of the processing step.

Anatomy, such as papillary muscle, is automatically detected (34) and/or detected in real-time. For automatic detection (34) of small anatomy, machine-learnt classification with spatial (32) and temporal (e.g., Markov) (34) constraints is used. For real-time detection, sparse machine-learnt detection (34) interleaved with optical flow tracking (38) is used.

A vision system of a vehicle includes a camera disposed at a vehicle and having a field of view exterior of the vehicle. The camera includes an imaging array having a plurality of photosensing elements arranged in a two dimensional array of rows and columns. The imaging array includes a plurality of sub-arrays comprising respective groupings of neighboring photosensing elements. An image processor is operable to perform a discrete cosine transformation of captured image data, and a Markov model compares at least one sub-array with a neighboring sub-array. The image processor is operable to adjust a classification of a sub-array responsive at least in part to the discrete cosine transformation and the Markov model.

Techniques for segmentation include determining an edge of voxels in a range associated with a target object. A center voxel is determined. Target size is determined based on the center voxel. In some embodiments, edges near the center are suppressed, markers are determined based on the center, and an initial boundary is determined using a watershed transform. Some embodiments include determining multiple rays originating at the center in 3D, and determining adjacent rays for each. In some embodiments, a 2D field of amplitudes is determined on a first dimension for distance along a ray and a second dimension for successive rays in order. An initial boundary is determined based on a path of minimum cost to connect each ray. In some embodiments, active contouring is performed using a novel term to refine the initial boundary. In some embodiments, boundaries of part-solid target objects are refined using Markov models.

The subject disclosure is directed towards a technology by which dynamic hand gestures are recognized by processing depth data, including in real-time. In an offline stage, a classifier is trained from feature values extracted from frames of depth data that are associated with intended hand gestures. In an online stage, a feature extractor extracts feature values from sensed depth data that corresponds to an unknown hand gesture. These feature values are input to the classifier as a feature vector to receive a recognition result of the unknown hand gesture. The technology may be used in real time, and may be robust to variations in lighting, hand orientation, and the user's gesturing speed and style.

The subject disclosure is directed towards a technology by which dynamic hand gestures are recognized by processing depth data, including in real-time. In an offline stage, a classifier is trained from feature values extracted from frames of depth data that are associated with intended hand gestures. In an online stage, a feature extractor extracts feature values from sensed depth data that corresponds to an unknown hand gesture. These feature values are input to the classifier as a feature vector to receive a recognition result of the unknown hand gesture. The technology may be used in real time, and may be robust to variations in lighting, hand orientation, and the user's gesturing speed and style.