Beijing Tusen Zhitu Technology Co., Ltd.

Tusimple, Inc.

Beijing Tusen Weilai Technology Co., Ltd.

Naiyan Wang

Zehao Huang

Xiaodi Hou

Pengfei Chen

Nan Yu

Hengchen Dai

Yuanqin Lu

Xuan Lin

Zhichao Li

Hongyang Li

Hao He

Zhenwei Shen

Ziqi Pang

Jianfu Zhang

Jiawei He

A method and device for training a neural network are disclosed. The method comprises: selecting, by a training device, a teacher network performing the same functions of a student network; and iteratively training the student network and obtaining a target network, through aligning distributions of features between a first middle layer and a second middle layer corresponding to the same training sample data, so as to transfer knowledge of features of a middle layer of the teacher network to the student network.

Method and device for student training networks with teacher networks

Embodiments of the present invention disclose a binocular image matching method, apparatus, device, and storage medium. The method comprises: performing target detection on a first image to obtain a first bounding box of a target in the first image; determining a second bounding box corresponding to the first bounding box in a second image; and obtaining a third bounding box of the target in the second image by regressing the second bounding box. The technical solutions realize accurate matching of targets in a binocular image without requiring performing target detection on both images in the binocular image, and then use a matching algorithm to match the targets detected in the two images, thereby greatly reducing the calculation overhead of target matching in the binocular image.

Binocular image matching method, device, and storage medium

The present disclosure provides a multiple target tracking method and apparatus, a calculating device and a storage medium, so as to solve the problem of inaccurate multiple target tracking in the prior art. The target tracking method comprises: obtaining a prediction box of a target in a current frame according to tracklets of one or more targets in historical frames; performing target detection on the current frame to obtain one or more detection boxes, wherein the detection boxes comprise a high-quality box and a medium-quality box; matching each prediction box with the detection box according to the similarity of the prediction box and the detection box; and in response to a prediction box being unmatched with the high-quality box but matched with the medium-quality box, determining that the target is in a tracking state in the current frame.

Multiple target tracking method and apparatus, calculating device and storage medium

Provided are a trajectory prediction method and device, a storage medium, and a computer program to avoid low accuracy and low reliability of a prediction result in conventional trajectory prediction methods. A trajectory prediction neural network acquires input current trajectory data and current map data of current environment when a moving subject moves in the current environment. The current trajectory data and the current map data are expressed as a current trajectory point set and a current map point set in a high-dimensional space. A global scene feature is extracted according to the current trajectory point set and the current map point set. The global scene feature has a trajectory feature and a map feature of the current environment. Multiple prediction trajectory point sets of the moving subject and a probability corresponding to each prediction trajectory point set are predicted and output according to the global scene feature.

Trajectory prediction method and device

An embodiment of the present disclosure discloses a training method for a multi-object tracking model and a multi-object tracking method. The multi-object tracking method comprises: constructing an object graph according to objects to be tracked in a current frame, wherein the vertexes of the object graph correspond to the objects to be tracked, and edge features of the edges between the two vertexes comprise an attribute relationship between the two vertexes; performing graph matching on the object graph and a tracklet graph to calculate matching scores between the object to be tracked and the tracked tracklet in the tracklet graph, wherein the vertexes of the tracklet graph correspond to tracked tracklets, and the edge features of the edges between the two vertexes comprise an attribute relationship between the two vertexes; and determining the matched tracklet of the object to be tracked according to the matching scores.

Training method for multi-object tracking model and multi-object tracking method

Disclosed are devices, systems and methods for processing an image. In one aspect a method includes receiving an image from a sensor array including an x-y array of pixels, each pixel in the x-y array of pixels having a value selected from one of three primary colors, based on a corresponding x-y value in a mask pattern. The method may further include generating a preprocessed image by performing preprocessing on the image. The method may further include performing perception on the preprocessed image to determine one or more outlines of physical objects.

Images for perception modules of autonomous vehicles

The present disclosure provides a method and an apparatus for creating an occupancy grid map, as well as a processing apparatus. The method includes: creating a current occupancy grid map based on a location of the vehicle and a previous occupancy grid map; and determining a current probability that each grid in the current occupancy grid map belongs to each of occupancy categories based on last environment perception information received from the sensors and updating an occupancy category to which each grid in the current occupancy grid map belongs based on the current probability that the grid belongs to each of the occupancy categories, in accordance with an asynchronous updating policy.

Method for creating occupancy grid map and processing apparatus

The present application relates to a method and an apparatus for detecting corner points of a lane line, an electronic device and a storage medium. The method comprises: obtaining a road image to be detected; inputting the road image into a pre-trained neural network model to obtain a heat map prediction result, wherein the heat map prediction result comprises a probability of each pixel belonging to a specific corner point category; determining a plurality of different categories of corner points in the road image according to the probability of each pixel belonging to a specific corner point category in the heat map prediction result; and grouping the plurality of different categories of corner points according to a predetermined rule to obtain at least one corner point group. The present application can solve the problem that the corner points of the lane lines cannot be accurately and quickly detected.

Method and apparatus for detecting corner points of lane lines, electronic device and storage medium

The present disclosure provides a method and an apparatus for binocular ranging, capable of achieving an improved accuracy of binocular ranging. The method includes: extracting features from a left image and a right image to obtain a left feature image and a right feature image; selecting a standard feature image and obtaining a cost volume of the standard feature image by applying a correlation calculation to the left feature image and the right feature image using a block matching algorithm; obtaining a confidence volume by normalizing computational costs of all disparity values in a disparity dimension for each pixel point in the cost volume; obtaining a confidence map by selecting a maximum value from confidence levels of all the disparity values in the disparity dimension for each pixel point in the confidence volume; obtaining a mask map by mapping each pixel point having a confidence level higher than a predetermined threshold in the confidence map to 1 and mapping each pixel point having a confidence level lower than or equal to the threshold in the confidence map to 0; obtaining a disparity map by calculating an argmax value for the confidence levels of all disparity values in the disparity dimension for each pixel point in the confidence volume; obtaining a target disparity map by multiplying the mask map with the disparity map; and estimating a distance based on the target disparity map.

Method and apparatus for binocular ranging

The present disclosure provides a method and an apparatus for semantic segmentation of an image, capable of solving the problem in the related art associated with low speed and inefficiency in semantic segmentation of images. The method includes: receiving the image; performing semantic segmentation on the image to obtain an initial semantic segmentation result; and inputting image information containing the initial semantic segmentation result to a pre-trained convolutional neural network for semantic segmentation post-processing, so as to obtain a final semantic segmentation result. With the solutions of the present disclosure, the initial semantic segmentation result can be post-processed using the convolutional neural network, such that the speed and efficiency of the semantic segmentation of the image can be improved.

Method and device for semantic segmentation of image

The present disclosure provides a method and an apparatus for training a neural network and a computer server. The method includes: selecting automatically input data for which processing by the neural network fails, to obtain a set of data to be annotated; annotating the set of data to be annotated to obtain a new set of annotated data; acquiring a set of newly added annotated data containing the new set of annotated data, and determining a union of the set of newly added annotated data and a set of training sample data for training the neural network in a previous period as a set of training sample data for a current period; and training the neural network iteratively based on the set of training sample data for the current period, to obtain a neural network trained in the current period.

Method and apparatus for training neural network and computer server

The present disclosure provides a method and an apparatus for sampling training data and a computer server. The method includes: inputting a video to a target detection model to obtain a detection result for each frame of image; inputting the detection results for all frames of images in the video to a target tracking model, to obtain a tracking result for each frame of image; and for each frame of image in the video: matching the detection result and the tracking result for the frame of image, and when the detection result and the tracking result for the frame of image are inconsistent with each other, determining the frame of image as a sample image to be marked, for which processing by the target detection model is not optimal.

Growing community of inventors

Beijing, China

Naiyan Wang