Road Surface Object Classification Using Vehicle Motion Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pothole detection methods face challenges such as limited feature space leading to lack of robustness, requiring large amounts of data, high computational complexity, and reliance on specific visual image data, which can result in poor classification results and increased energy costs.
Innovation Solution
A method using a computer-implemented approach that classifies road surface objects by analyzing horizontal and vertical vehicle motions through a fleet of vehicles, clustering data points, and employing a convolutional neural network (CNN) to process scatter plots of speed and vertical displacement amplitudes, enabling backend processing and digital map updates for driver warnings and automated driving functions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine learning approaches with handcrafted features are used for road surface object classification, then domain expertise can be incorporated, but the feature engineering process becomes complex and requires extensive expert knowledge
Solution Approach 1:
The patent replaces the manual mechanical process of handcrafting features with an automated neural network-based feature extraction system. The neural network automatically learns relevant features from raw sensor data, eliminating the need for expert-driven feature engineering while maintaining or improving classification robustness.
Solution Approach 2:
The neural network performs self-service by automatically extracting features from raw data without human intervention. The system trains the network to identify relevant patterns in sensor measurements, allowing the model to autonomously determine which features are important for road surface object classification.
2Extent of automation
If deep learning classification approaches are used, then automation is improved, but computational complexity and energy consumption increase
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network offline using extensive computational resources. Once trained, the model can be deployed in vehicles where it performs classification with minimal real-time computational requirements, shifting the energy burden from operational to training phase.
Solution Approach 2:
The patent extracts the computationally intensive training process from the real-time classification system. By separating offline model development from online inference, the system achieves high automation in vehicle operation while minimizing energy consumption during actual driving.
3Loss of information
If optical data collection methods are used for pothole detection, then visual information can be obtained, but the system requires advanced sensors and increases device complexity
Solution Approach 1:
The patent replaces optical sensing systems (cameras, LIDAR) with mechanical vibration sensors that are already present in vehicles. By analyzing vibration patterns from the suspension system, the system detects road surface conditions without requiring additional complex optical hardware.
Solution Approach 2:
The patent makes existing vibration sensors serve multiple functions: they continue to monitor vehicle dynamics for suspension control while simultaneously detecting road surface objects like potholes. This multi-functionality eliminates the need for dedicated optical sensing systems.
4Measurement precision
If more features are used for classification, then measurement precision improves, but the quantity of data and computational requirements increase
Solution Approach 1:
The patent changes the parameter representation by transforming raw sensor data into frequency domain characteristics through Fourier analysis. This transformation extracts essential vibration patterns into compact spectral features, maintaining classification precision while reducing data volume and computational requirements.
Solution Approach 2:
The patent extracts only the most relevant frequency components from the vibration signal using spectral analysis. By focusing on specific frequency ranges that correspond to road surface object characteristics, the system achieves precise classification without processing the entire raw signal spectrum.
Data Source
AI summary
A road surface is classified by providing a set of data points that is attributable to a same road surface object. Each data point specifies a first variable and a second variable. For each data point, the first variable characterizes a horizontal motion exhibited by a vehicle when driving over the road surface object and the second variable characterizes a vertical motion exhibited by said vehicle when driving over the road surface object. The set of data points are classified using an artificial neural network with regard to a relevance of the road surface object for a driver warning function or an automated driving function.