Unified Trajectory Prediction Network for Mixed Road Users
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural networks (DNNs) for autonomous driving systems are inefficient as they require separate models for different types of objects, such as vehicles, pedestrians, and cyclists, and lack resource sharing, making them unsuitable for predicting future trajectories of various objects effectively.
Innovation Solution
A deep neural network structure that includes a shared information generation module and a future trajectory prediction module, utilizing a convolutional neural network (CNN) and long short-term memory (LSTM) to generate driving environment feature maps and motion feature vectors, allowing for the prediction of future trajectories of various types of objects by transforming data into object-centered coordinates and using variational auto-encoders (VAEs) for trajectory generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate DNN models are used for different object types (vehicle, pedestrian, cyclist), then prediction accuracy for each object type can be optimized, but system complexity increases and resource sharing becomes impossible
Solution Approach 1:
The patent implements a universal DNN model that can process and predict trajectories for multiple object types (vehicles, pedestrians, cyclists) using a single unified architecture. The model accepts object type as an input parameter and adapts its processing accordingly, eliminating the need for separate specialized models for each object type while maintaining prediction accuracy.
Solution Approach 2:
The patent segments the trajectory prediction task by object type through internal processing stages within the unified model. Different object types are handled through type-specific processing paths or attention mechanisms within the same model framework, allowing specialized treatment for each object type without requiring separate model instances.
2Reliability
If multiple separate DNN models are maintained for different object types, then specialized prediction capabilities are achieved, but resource sharing between models becomes impossible and computational efficiency decreases
Solution Approach 1:
The patent merges multiple object-type-specific prediction models into a single unified DNN model. This consolidation allows the system to share computational resources, including model parameters, computation graphs, and memory, across different object types. The unified model processes multiple object types sequentially or through multi-task learning mechanisms, improving computational efficiency while maintaining reliable predictions for each object type.
3Productivity
If a unified DNN model is used for all object types, then resource sharing and computational efficiency improve, but the model must handle diverse object characteristics which may reduce prediction accuracy
Solution Approach 1:
The patent applies local quality by incorporating object-type-specific processing mechanisms within the unified model. Different object types receive customized processing treatments through type-aware attention mechanisms, object-type-specific feature extraction layers, or conditional processing paths. This allows the model to maintain high prediction accuracy for each object type while benefiting from the resource efficiency of a unified architecture.
Solution Approach 2:
The patent implements dynamic adaptability within the unified model by using object type information to dynamically adjust processing parameters, attention weights, or feature extraction strategies. The model can switch between different processing modes or emphasize different features based on the input object type, allowing it to optimize predictions for vehicles, pedestrians, and cyclists differently within the same computational framework.
Data Source
AI summary
The present disclosure relates to an apparatus and a method for predicting future trajectories of various types of objects using an artificial neural network trained by a method for training an artificial neural network to predict future trajectories of various types of moving objects for autonomous driving. The apparatus for predicting future trajectories includes a shared information generation module configured to: collect location information of one or more objects around an autonomous vehicle for a predetermined time, generate past movement trajectories for the one or more objects based on the location information, and generate a driving environment feature map for the autonomous vehicle based on road information around the autonomous vehicle and the past movement trajectories; and a future trajectory prediction module configured to generate future trajectories for the one or more objects based on the past movement trajectories and the driving environment feature map.