Vehicle Collision Prediction Using Top-View CNN Risk Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Advanced driver assistance systems (ADAS) face challenges in handling various collision situations during autonomous driving, as they often struggle to predict collisions with nearby vehicles and pedestrians effectively, leading to difficulties in implementing flexible collision avoidance strategies.
Innovation Solution
A vehicle equipped with multiple sensors and a processor using an artificial neural network model, specifically a pre-trained convolutional neural network (CNN), generates a top view image to determine the degree of collision risk and collision mode, allowing for proactive collision avoidance and passive safety system deployment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the vehicle only predicts whether a collision will occur based on sensing information, then the prediction process is simple, but it is difficult to deal flexibly with various collision situations
Solution Approach 1:
The collision prediction is divided into multiple dimensions: collision risk determination (binary classification) and collision mode determination (multi-class classification including front collision, rear collision, side collision, and no collision). This segmentation allows the system to handle various collision situations flexibly while maintaining manageable complexity through modular processing
Solution Approach 2:
The system transitions from simple 2D sensor data to a top-view image representation that adds spatial dimensionality. The top-view image integrates position, angle, moving path, and collision risk information into a comprehensive visual format, enabling the neural network to analyze collision scenarios from multiple spatial perspectives simultaneously
2Measurement precision
If multiple sensors are used to obtain state information on nearby vehicles, then collision prediction accuracy is improved, but system complexity and computational load increase
Solution Approach 1:
Multiple sensors (cameras, LIDAR, radar) are merged into a unified top-view image representation. The sensor fusion occurs at the image level rather than raw data level, where position, angle, moving path, and collision risk information from different sensors are integrated into a single comprehensive visual input for the neural network, improving accuracy while managing complexity
Solution Approach 2:
The top-view image serves as an intermediary representation between raw sensor data and the neural network analysis. This intermediate format consolidates complex multi-sensor information into a standardized visual structure, facilitating accurate collision prediction while simplifying the processing pipeline
3Measurement precision
If a top view image is generated to represent moving path, position, angle, and collision risk, then collision mode determination accuracy is improved, but image generation and processing time increase
Solution Approach 1:
The top-view image is generated in advance using pre-calculated parameters such as collision risk indices, moving paths, and vehicle states. By preparing this comprehensive visual representation before neural network analysis, the system enables accurate collision mode determination while optimizing processing time through efficient pre-computation of key parameters
Data Source
AI summary
A vehicle includes: a plurality of sensors configured to obtain state information on a nearby vehicle; and a processor that is operatively connected to the plurality of sensors. The processor is configured to: determine a degree of risk of collision between the vehicle and the nearby vehicle based on the state information on the nearby vehicle; generate a top view image based on at least one of the state information or the degree of risk of collision; and determine a collision mode which indicates whether the vehicle collides with at least on of the nearby vehicle or a predicted collision portion through an artificial neural network model using the top view image as an input.