Monocular Vehicle Ranging With Depth Fusion for Intelligent Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ranging methods for intelligent driving, such as those using binocular vision systems, face limitations in measurement accuracy, application flexibility, and high cost due to the need for high-computing capabilities and synchronized cameras with parallel optical axes.
Innovation Solution
A monocular ranging method utilizing a vehicle-mounted monocular camera, involving a two-stage convolutional neural network for instance segmentation and depth estimation, followed by clustering and geometric ranging models to compute distances, with fusion computing to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If binocular vision systems are used for ranging, then measurement accuracy is improved, but device complexity and cost increase due to requiring two synchronized cameras with parallel optical axes and high computing capabilities
Solution Approach 1:
The patent extracts the ranging function from a complex binocular system and implements it using a single camera combined with monocular depth estimation neural networks and geometric models, eliminating the need for multiple synchronized cameras while maintaining measurement capability
Solution Approach 2:
The patent replaces the mechanical/optical binocular ranging system with a computational approach using monocular vision and neural networks, substituting physical complexity with algorithmic processing to achieve ranging functionality
2Measurement precision
If binocular vision systems are used for ranging, then measurement accuracy is improved, but computing power requirements increase due to parallax-based depth estimation
Solution Approach 1:
The patent replaces computationally intensive binocular parallax calculations with monocular depth estimation neural networks that process single images, significantly reducing computing power requirements while maintaining ranging accuracy through learned depth priors and geometric constraints
3Measurement precision
If binocular vision systems are used for ranging, then measurement accuracy is improved, but cost increases due to requiring two cameras and high computing capabilities
Solution Approach 1:
The patent uses a single, lower-cost monocular camera instead of expensive binocular camera systems, accepting that the simpler sensor can achieve accurate ranging through computational methods, thereby reducing overall system cost while maintaining functionality
4Measurement precision
If binocular vision systems are used for ranging, then measurement accuracy is improved, but application flexibility decreases due to requiring synchronized acquisition and consistent exposure parameters
Solution Approach 1:
The patent extracts the ranging capability from the constraints of binocular synchronization and exposure matching, enabling independent camera operation and broader application flexibility through monocular depth estimation that does not require coordinated multi-sensor operation
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to the technical field of intelligent driving, and in particular, to a monocular ranging method for intelligent driving. The monocular ranging method for intelligent driving includes: acquiring image information and inputting the image information into a first convolutional neural network model for detection to obtain a segmentation image and a depth map, the segmentation image containing detection boxes of targets; extracting a rectangular box corresponding to the detection box of each target, segmenting the segmentation image according to the rectangular box to obtain a region of interest (ROI) image corresponding to the target, and inputting the ROI image into a second convolutional neural network model for detection to obtain information of the target; processing the segmentation image and the depth map and computing a first distance by using a clustering computation model; computing a second distance by using a geometric ranging model based on the information of the target; and performing fusion computing on the first distance and the second distance to finally obtain a distance to the target in real space. The monocular ranging method for intelligent driving of the present disclosure improves the measurement accuracy, is not limited in use, and has a low cost.