Vision-Based Collision Avoidance Path Planning for Vehicles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing advanced driver assistance systems (ADAS) for collision avoidance rely on noisy data from distance sensors like radar and lidar, which are costly and difficult to produce in large quantities, and struggle to accurately recognize pedestrians, while camera sensors provide insufficient detail for collision avoidance.
Innovation Solution
A vision-based collision avoidance system using semantic segmentation, a deep learning vision recognition technology, separates obstacles, travelable regions, and non-travelable regions in a vehicle's field of view, generating an optimal avoidance path without relying on active sensors, thereby reducing production costs and preventing false-positive targets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If radar or lidar sensors are used for collision avoidance, then measurement precision is improved, but manufacturing cost increases
Solution Approach 1:
The patent replaces expensive radar and lidar sensors with inexpensive camera sensors. The camera, being a low-cost component, is used to capture visual information for collision avoidance, eliminating the need for costly active sensors while maintaining functional capability through image processing and deep learning algorithms.
Solution Approach 2:
The patent substitutes active sensing mechanisms (radar, lidar) with passive optical sensing (camera). Instead of emitting electromagnetic waves or light to detect objects, the system uses a camera to capture visual scenes, then processes images through neural networks to identify obstacles and generate avoidance paths, replacing mechanical/electromagnetic sensing with optical sensing and computational processing.
2Measurement precision
If radar sensor is used for collision avoidance, then relative distance and speed are detected accurately, but false-positive targets occur
Solution Approach 1:
The patent replaces radar sensors that generate false positives with camera sensors combined with deep learning image processing. The visual recognition system processes image data through neural networks to identify actual obstacles, eliminating false-positive detections while maintaining accurate distance and speed measurement capabilities through computer vision techniques.
Solution Approach 2:
The patent introduces deep learning image processing as an intermediary between the camera sensor and collision avoidance decision-making. The neural network processes raw image data, filters out false detections, and identifies genuine obstacles, serving as a mediator that eliminates false positives while preserving accurate target detection.
3Ease of manufacture
If camera sensor is used for collision avoidance, then production cost is reduced, but measurement precision deteriorates
Solution Approach 1:
The patent substitutes simple camera sensing with an enhanced visual processing system using deep learning. By replacing traditional image processing with neural networks, the system achieves radar-level detection accuracy using only passive optical sensing, eliminating the need for expensive active sensors while maintaining high measurement precision through computational intelligence.
Solution Approach 2:
The patent transforms the camera system from a simple imaging device into a precision measurement tool by changing the processing parameters through deep learning. The neural network extracts multiple parameters (distance, speed, obstacle type, trajectory) from image data, enabling the low-cost camera to achieve high measurement precision comparable to expensive sensors through advanced algorithmic processing.
Data Source
Figure 1
Figure 2(a)~2(c)
Figure 3
AI summary
A system for avoiding a collision based on a vision includes a camera for filming a region ahead of a vehicle, a path generating device that sets a travelable region where the vehicle may travel by avoiding an obstacle in a vision of the region ahead of the vehicle filmed by the camera, and generates one or more travel paths in the travelable region, and a motion controller that selects one optimal path from the travel paths to perform a longitudinal control or a lateral control of the vehicle such that the vehicle travels along the optimal path.