HD Map Feature Fusion for Vehicle Behavior Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving technologies fail to accurately predict the behavior of objects in the surrounding environment of an autonomous vehicle due to the lack of comprehensive representation, particularly without utilizing information from high definition maps.
Innovation Solution
A computer-implemented method that extracts features from high definition maps using a convolutional neural network and concatenates them with handcrafted features of target objects, allowing the network to predict the behavior of the objects by providing a comprehensive representation of the driving environment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If existing approaches focus only on target object features, then the prediction model is simpler, but the representation of surrounding environment is incomplete
Solution Approach 1:
The patent merges handcrafted features (geometric, semantic, topological) with deep learning features from CNNs to create a comprehensive environmental representation. This combination integrates multiple feature sources to reduce information loss while managing system complexity through structured feature fusion.
Solution Approach 2:
The patent creates a composite feature representation by combining different types of features (handcrafted geometric/semantic/topological features with learned deep features) into a unified feature vector, analogous to composite materials that combine different substances to achieve superior properties.
2Measurement precision
If comprehensive environmental representation is achieved by integrating HD map features, then prediction accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the feature extraction process into distinct modules: handcrafted feature extraction (geometric, semantic, topological), deep feature extraction from HD maps, and feature fusion. This segmentation allows each module to be optimized independently while achieving comprehensive representation for accurate prediction.
Solution Approach 2:
The patent performs preliminary feature extraction and processing before the main prediction task. Handcrafted features are pre-computed from HD map data and target object data, and deep features are extracted in advance, allowing the final prediction model to work with pre-processed comprehensive features rather than raw data.
3Reliability
If handcrafted features are combined with deep learning features, then the comprehensive representation is achieved, but the system complexity increases
Solution Approach 1:
The patent transforms various input parameters (geometric properties, semantic attributes, topological relationships) into standardized feature vectors with consistent dimensions and formats. This parameter transformation allows diverse handcrafted features to be seamlessly integrated with deep learning features, achieving reliable representation while managing complexity through uniform parameter handling.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
In an embodiment, a method for representing a surrounding environment of an ego autonomous driving vehicle (ADV) is described. The method represents the surrounding environment using a first set of features from a definition (HD) map and a second set of features from a target object in the surrounding environment. The first set of features are extracted from the high definition map using a convolutional neural network (CNN), and the second set of features are handcrafted features from the target object during a predetermined number of past driving cycles of the ego ADV. The first set of features and the second set of features are concatenated and provided to a number of fully connected layers of the CNN to predict behaviors of the target object. In one embodiment, the operations in the method can be repeated for each driving cycle of the ego ADV.