Aptiv Technologies Ag

Aptiv Technologies Limited

Kanishka Tyagi

Narbik Manukian

Shan Zhang

Yihang Zhang

John Kirkwood

Nianxia Cao

Xiaohui Wang

Sanling Song

Steven Shaw

Kaveh Ahmadi

This document describes techniques and systems for machine-learning-based super resolution of radar data. A low-resolution radar image can be used as input to train a model for super resolution of radar data. A higher-resolution radar image, generated by an effective, but costly in terms of computing resources, traditional super resolution method, and the higher-resolution image can serve as ground truth for training the model. The resulting trained model may generate a high-resolution sensor image that closely approximates the image generated by the traditional method. Because this trained model needs only to be executed in feed-forward mode in the inference stage, it may be suited for real-time applications. Additionally, if low-level radar data is used as input for training the model, the model may be trained with more comprehensive information than can be obtained in detection level radar data.

Machine-learning-based super resolution of radar data

This document describes systems and techniques related to deep association for sensor fusion. For example, a model trained using deep machine learning techniques, may be used to generate an association score matrix that includes probabilities that tracks from different types of sensors are related to the same objects. This model may be trained using a convolutional recurrent neural network and include constraints not included in other training techniques. Focal loss can be used during training to compensate for imbalanced data samples and address difficult cases, and data expansion techniques can be used to increase the multi-sensor data space. Simple thresholding techniques can be applied to the association score matrix to generate an assignment matrix that indicates whether tracks from one sensor and tracks from another sensor match. In this manner, the track association process may be more accurate than current sensor fusion techniques, and vehicle safety may be increased.

Deep association for sensor fusion

This document describes techniques and systems for stationary object detection and classification based on low-level radar data. Raw electromagnetic signals reflected off stationary objects and received by a radar system may be preprocessed to produce low-level spectrum data in the form of range-Doppler maps that retain all or nearly all the data present in the raw electromagnetic signals. The preprocessing may also filter non-stationary range-Doppler bins. The remaining low-level spectrum data represents stationary objects present in a field-of-view (FOV) of the radar system. The low-level spectrum data representing stationary objects can be fed to an end-to-end deep convolutional detection and classification network that is trained to classify and provide object bounding boxes for the stationary objects. The outputted classifications and bounding boxes related to the stationary objects may be provided to other driving systems to improve their functionality resulting in a safer driving experience.

Stationary object detection and classification based on low-level radar data

This document describes techniques and systems related to a radar system using a machine-learned model for stationary object detection. The radar system includes a processor that can receive radar data as time-series frames associated with electromagnetic (EM) energy. The processor uses the radar data to generate a range-time map of the EM energy that is input to a machine-learned model. The machine-learned model can receive as inputs extracted features corresponding to the stationary objects from the range-time map for multiple range bins at each of the time-series frames. In this way, the described radar system and techniques can accurately detect stationary objects of various sizes and extract critical features corresponding to the stationary objects.

Radar system using a machine-learned model for stationary object detection

This document describes techniques and systems for fuzzy labeling of low-level electromagnetic sensor data. Sensor data in the form of an energy spectrum is obtained and the points within an estimated geographic boundary of a scatterer represented by the smear is labeled with a value of one. The remaining points of the energy spectrum are labeled with values between zero and one with the values decreasing the further away each respective remaining point is from the geographic boundary. The fuzzy labeling process may harness more in-depth information available from the distribution of the energy in the energy spectrum. A model can be trained to efficiently label an energy spectrum map in this manner. This may result in lower computational costs than other labeling methods. Additionally, false detections by the sensor may be reduced resulting in more accurate detection and tracking of objects.

Fuzzy labeling of low-level electromagnetic sensor data

Techniques and apparatuses are described that implement object classification using low-level radar data. In particular, a radar system extracts features of a detected object based on low-level data. The radar system analyzes these features using machine learning to determine an object class associated with the detected object. By relying on low-level data, the radar system is able to extract additional information regarding the distribution of energy across range, range rate, azimuth, or elevation, which is not available in detection-level data. With the use of machine learning, the object can be classified quickly (e.g., within a single frame or observation), thereby enabling sufficient time for the autonomous-driving logic to initiate an appropriate action based on the object's class. Furthermore, this classification can be performed without the use of information from other sensors.

Growing community of inventors

Agoura Hills, CA, United States of America

Kanishka Tyagi