Fca US LLC

Dalong Li

Stephen W Horton

Neil R Garbacik

Rohit S Paranjpe

Andrew Chen

Ivan Roman

Alex M Smith

Zijian Wang

Ayyoub Rezaeian

Benjamin J Chappell

Vehicle perception techniques include applying a 3D DNN to a set of inputs to generate 3D detection results including a set of 3D objects, transforming the set of 3D objects onto a set of images as a first set of 2D bounding boxes, applying a 2D DNN to the set of images to generate 2D detection results including a second set of 2D bounding boxes, calculating mean average precision (mAP) values based on a comparison between the first and second sets of 2D bounding boxes, identifying a set or corner cases based on the calculated mAP values, and re-training or updating the 3D DNN using the identified set of corner cases, wherein a performance of the 3D DNN is thereby increased without the use of expensive additional manually and/or automatically annotated training datasets.

Automatic evaluation of three-dimensional vehicle perception using two-dimensional deep neural networks

Vehicle perception techniques include obtaining a training dataset represented by N training histograms, in an image feature space, corresponding to N training images, K-means clustering the N training histograms to determine K clusters with respective K respective cluster centers, wherein K and N are integers greater than or equal to one and K is less than or equal to N, comparing the N training histograms to their respective K cluster centers to determine maximum in-class distances for each of K clusters, applying a deep neural network (DNN) to input images of the set of inputs to output detected/classified objects with respective confidence scores, obtaining adjusted confidence scores by adjusting the confidence scores output by the DNN based on distance ratios of (i) minimal distances of input histograms representing the input images to the K cluster centers and (ii) the respective maximum in-class.

Vehicle perception by adjusting deep neural network confidence valves based on k-means clustering

An object detection and classification verification system for a vehicle includes a projection system configured to project a three-dimensional (3D) scene pre-captured at a known distance and comprising at least one known object onto a surface in front of the vehicle a controller configured to verify a performance of an object detection and classification routine by performing the object detection and classification routine on the projected 3D scene to generate a set of results, comparing the set of results to a set of expected results associated with the projected 3D scene, and based on the comparing, determining whether the performance of the object detection and classification routine satisfies a predetermined threshold metric.

Techniques to automatically verify object detection, classification, and depth for automated driving systems

Systems and methods for testing a machine learning algorithm or technique of an autonomous driving feature of a vehicle utilize a sensor system configured to capture input data representative of an environment external to the vehicle and a controller configured to receive the input data from the sensor system and perform a testing procedure for the autonomous driving feature that includes inserting known input data into a target portion of the input data to obtain modified input data, processing the modified input data according to the autonomous driving feature to obtain output data, and determining an accuracy of the autonomous driving features based on a comparison between the output data and the known input data.

Functionally safe rationalization check for autonomous vehicle machine learning algorithms

Vehicle center of gravity (CoG) height and mass estimation techniques utilize a light detection and ranging (LIDAR) sensor configured to emit light pulses and capture reflected light pulses that collectively form LIDAR point cloud data and a controller configured to estimate the CoG height and the mass of the vehicle during a steady-state operating condition of the vehicle by processing the LIDAR point cloud data to identify a ground plane, identifying a height difference between (i) a nominal distance from the LIDAR sensor to the ground plane and (ii) an estimated distance from the LIDAR sensor to the ground plane using the processed LIDAR point cloud data, estimating the vehicle CoG height as a difference between (i) a nominal vehicle CoG height and the height difference, and estimating the vehicle mass based on one of (i) vehicle CoG metrics and (ii) dampening metrics of a suspension of the vehicle.

Vehicle center of gravity height detection and vehicle mass detection using light detection and ranging point cloud data

Techniques for training multiple resolution deep neural networks (DNNs) for vehicle autonomous driving comprise obtaining a training dataset for training a plurality of DNNs for an autonomous driving feature of the vehicle, sub-sampling the training dataset to obtain a plurality of training datasets comprising the training dataset and one or more sub-sampled datasets each having a different resolution than a remainder of the plurality of training datasets, training the plurality of DNNs using the plurality of training datasets, respectively, determining a plurality of outputs for the autonomous driving feature using the plurality of trained DNNs and the input data, receiving input data for the autonomous driving feature captured by a sensor device, and determining a best output for the autonomous driving feature using the plurality of outputs.

Multiple resolution deep neural networks for vehicle autonomous driving systems

A semi-automatic three-dimensional light detection and ranging (LIDAR) point cloud data annotation system and method for autonomous driving of a vehicle involve filtering 3D LIDAR point cloud and normalizing the filtered 3D LIDAR point cloud data relative to the vehicle to obtain normalized 3D LIDAR point cloud data, quantizing the normalized 3D LIDAR point cloud data by dividing it into a set of 3D voxels, projecting the set of 3D voxels to a 2D birdview, identifying a possible object by applying clustering to the 2D birdview projection, obtaining an annotated 2D birdview projection including annotations by a human annotator via the annotation system regarding whether the bounding box corresponds to a confirmed object and a type of the confirmed object, and converting the annotated 2D birdview projection to back into annotated 3D LIDAR point cloud data.

Semi-automatic LIDAR annotation system for autonomous driving

An advanced driver assistance system (ADAS) and method for a vehicle utilize a light detection and ranging (LIDAR) system configured to emit laser light pulses and capture reflected laser light pulses collectively forming three-dimensional (3D) LIDAR point cloud data and a controller configured to receive the 3D LIDAR point cloud data, convert the 3D LIDAR point cloud data to a two-dimensional (2D) birdview projection, detect a set of lines in the 2D birdview projection, filter the detected set of lines to remove lines having features that are not indicative of traffic signs to obtain a filtered set of lines, and detect one or more traffic signs using the filtered set of lines.

Traffic sign detection from filtered birdview projection of LIDAR point clouds

An advanced driver assistance system (ADAS) and method for a vehicle utilize a light detection and ranging (LIDAR) system configured to emit laser light pulses and capture reflected laser light pulses collectively forming three-dimensional (3D) LIDAR point cloud data and a controller configured to receive the 3D LIDAR point cloud data, convert the 3D LIDAR point cloud data to a two-dimensional (2D) birdview projection, obtain a template image for object detection, the template image being representative of a specific object, blur the 2D birdview projection and the template image to obtain a blurred 2D birdview projection and a blurred template image, and detect the specific object by matching a portion of the blurred 2D birdview projection to the blurred template image.

Tracking objects in LIDAR point clouds with enhanced template matching

An advanced driver assistance system (ADAS) and corresponding method for a vehicle utilize a camera system configured to capture an image and a controller configured to receive the captured image, detect an object in the captured image using a simple neural network model, track the detected object using a tracking technique to obtain a tracked position, project a trajectory of the detected object using a trajectory projection technique to obtain a predicted position, determine a most likely position of the detected object based on at least one of the tracked and predicted positions, generate a two-dimensional (2D) birdview projection illustrating the detected object according to its determined most likely position, and control at least one ADAS feature of the vehicle using the generated 2D birdview projection.

Techniques for using a simple neural network model and standard camera for image detection in autonomous driving

An advanced driver assistance system (ADAS) and method for a vehicle utilize a light detection and ranging (LIDAR) system configured to emit laser light pulses and capture reflected laser light pulses collectively forming three-dimensional (3D) LIDAR point cloud data and a controller configured to receive the 3D LIDAR point cloud data divide the 3D LIDAR point cloud data into a plurality of cells corresponding to distinct regions surrounding the vehicle, generate a histogram comprising a calculated height difference between a maximum height and a minimum height in the 3D LIDAR point cloud data for each cell of the plurality of cells, and using the histogram, perform at least one of adaptive ground removal from the 3D LIDAR point cloud data and traffic level recognition.

Growing community of inventors

Troy, MI, United States of America

Dalong Li