Distance to obstacle detection in autonomous machine applications

Abstract

In various examples, a deep neural network (DNN) is trained to accurately predict, in deployment, distances to objects and obstacles using image data alone. The DNN may be trained with ground truth data that is generated and encoded using sensor data from any number of depth predicting sensors, such as, without limitation, RADAR sensors, LIDAR sensors, and / or SONAR sensors. Camera adaptation algorithms may be used in various embodiments to adapt the DNN for use with image data generated by cameras with varying parameters-such as varying fields of view. In some examples, a post-processing safety bounds operation may be executed on the predictions of the DNN to ensure that the predictions fall within a safety-permissible range.

Description

Background technique [0001] Being able to correctly detect the distance between a vehicle (such as an autonomous or semi-autonomous vehicle) and objects or obstacles in the environment is critical to the safe operation of the vehicle. For example, accurate distances to obstacles based on sensor data estimates are central to longitudinal control tasks such as automatic cruise control (ACC) and automatic emergency braking (AEB) as well as lateral control tasks such as safety checks for lane changes and safe lane change execution . [0002] Conventional methods of calculating distances to objects or obstacles in a vehicle's environment rely on the assumption that the ground plane or earth is flat. Based on this assumption, two-dimensional (2D) information sources, such as 2D images, can be used to model three-dimensional (3D) information. For example, because the ground plane is assumed to be flat, conventional systems also assume that the bottom of the two-dimensional bounding...

AI Technical Summary

Problems solved by technology

[0003] However, these traditional methods suffer when the actual road surface defining the actual ground plane is curved or otherwise uneven

For example, driving curves in the surface leads to inaccurate predictions—e.g., overestimation or underestimation—of distances relative to objects or obstacles in the environment when the imposed assumption that the ground plane is flat is in fact not

In either case, inaccurate distance estimation has immediate negative consequences on various operations of the vehicle, thereby potentially compromising the safety, performance and reliability of lateral and longitudinal control or warning-related driving characteristics

As an example, an underestimated distance could result in a failure to engage ACC and, even more critically, the AEB function to prevent a potential collision

Conversely, an overestimated distance may result in the inability to activate the ACC or AEB functions when not required, causing potential discomfort or injury to the occupants, while also reducing occupant confidence in the vehicle's ability to perform safely

[0004] Another disadvantage of traditional systems is the reliance on unification between the cameras used during training and those used in deployment.

For example, because deep neural networks (DNNs) can learn from scaling of objects and surroundings, limitations arise when the image data used during deployment is generated by cameras with different parameters than those used in training

For example, objects in the training image data can be scaled differently than those in the deployed image data, leading to inaccurate predictions by the DNN relative to the distances of objects or obstacles

These inaccurate predictions can lead to similar issues above regarding the accuracy and effectiveness of the vehicle's various driving tasks

Images