Traffic Primitive Indexing for Heterogeneous Driving Datasets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing traffic datasets are heterogeneous and difficult to combine due to differences in data collection platforms and sensor formats, making it challenging to train or test learning-based algorithms for autonomous vehicles.
Innovation Solution
A system that unifies multiple traffic datasets by combining records into a single dataset ordered by timestamps, using Bayesian learning methods like hidden Markov models to identify traffic primitives, and creating a traffic primitive index for querying specific records, allowing for easier training or testing of learning-based algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If multiple heterogeneous datasets are combined, then the quantity of usable traffic data increases, but the complexity of data integration increases due to differences in data collection platforms and sensor formats
Solution Approach 1:
The system segments the heterogeneous dataset into multiple tables, with each table corresponding to a specific sensor type (e.g., image sensors, LiDAR sensors, RADAR, GPS, IMU). This segmentation allows each sensor's data to be organized and processed independently according to its specific format and characteristics, reducing the complexity of integrating heterogeneous data while preserving the ability to access all data through a unified interface
Solution Approach 2:
The patent introduces a primitive index as an intermediary layer between the heterogeneous sensor data and the user queries. The primitive index contains segmented records with standardized fields that map to different sensor types, acting as a mediator that translates diverse sensor formats into a unified structure that can be efficiently queried without requiring complex real-time integration of the underlying heterogeneous data
2Measurement precision
If extensive post-processing is performed to extract interested information from chronologically logged datasets, then the precision of extracted information improves, but the time and effort required increases
Solution Approach 1:
The system performs preliminary action by pre-segmenting the chronological records into traffic primitives and organizing them into sensor-specific tables during data ingestion. This pre-processing creates a structured format with standardized fields that directly represent meaningful traffic events, eliminating the need for extensive post-processing to extract interested information and allowing users to query pre-organized data structures
Solution Approach 2:
The patent transforms the data parameters by converting raw chronological sensor logs into segmented records with standardized fields that represent traffic primitives. This parameter transformation changes the data from a continuous chronological stream into discrete, categorized events with explicit attributes, making the extracted information immediately usable without additional processing time
Data Source
AI summary
In one embodiment, example systems and methods related to a manner of unifying heterogeneous datasets are provided. Multiple heterogeneous datasets containing traffic or driving data are collected. The records of the datasets are combined, and the records in the combined dataset are ordered into a plurality of time series based on timestamps associated with each record. A Bayesian learning method, such as hidden Markov models, is used to identify traffic primitives in the datasets. Each traffic primitive may include several consecutive records in the combined dataset and may correspond to particular driving actions such as turning left or right, stopping, accelerating, etc. The traffic primitives are used to create a traffic primitive index that can be queried by users or researchers for specific records. These records can be used to train or test one or more learning-based algorithms. In addition, the combined dataset can be further divided into tables corresponding to particular sensors, allowing the users or researchers to query for specific traffic primitive and sensor combinations.