Probabilistic Neural Network For Predicting Hidden Context Of Traffic Entities For Autonomous Vehicles

What is this patent about?

’579 is related to the field of autonomous vehicle navigation and, more specifically, to systems that improve autonomous navigation by predicting the behavior of other traffic participants. Current autonomous systems often struggle to accurately predict the actions of pedestrians or cyclists, leading to erratic vehicle behavior. The patent addresses this problem by incorporating a prediction of the 'hidden context' of other traffic entities, such as their intentions or awareness of the autonomous vehicle.

The underlying idea behind ’579 is to use a probabilistic neural network to infer the hidden context of traffic entities from camera images and to use the uncertainty of this inference to guide the autonomous vehicle's navigation. The neural network is trained using data from human observers who have viewed similar traffic scenarios and provided their assessments of the likely behavior of the traffic entities. This allows the autonomous vehicle to make more human-like judgments about the intentions of other road users.

The claims of ’579 focus on a method, a storage medium, and a computer system for training and executing a probabilistic neural network to predict the hidden context of traffic entities. The independent claims cover receiving an image of traffic, generating a feature vector with statistical distributions, generating an output representing the hidden context (as likelihoods of user responses), determining a measure of uncertainty for these likelihoods, and navigating the autonomous vehicle based on this uncertainty.

In practice, the system captures images of traffic using a camera mounted on the autonomous vehicle. The probabilistic neural network processes these images to estimate the hidden context of nearby pedestrians, cyclists, or other vehicles. For example, the system might predict the likelihood that a pedestrian intends to cross the street and how confident it is in that prediction. The uncertainty measure is crucial because it allows the autonomous vehicle to adjust its behavior based on the reliability of the prediction. A high uncertainty might lead to a more cautious approach, while a low uncertainty allows for more assertive navigation.

The key differentiation from prior approaches lies in the use of a probabilistic neural network and the explicit consideration of uncertainty. Traditional systems often rely on kinematic models that extrapolate future motion based on current movement, which can be unreliable for predicting human behavior. By incorporating human judgment and quantifying the uncertainty in its predictions, ’579 enables the autonomous vehicle to make more robust and human-like navigation decisions , especially in complex and unpredictable traffic scenarios. The system generates a statistical distribution of possible outcomes, rather than a single point estimate, allowing for a more nuanced and safer navigation strategy.