Road Object Detection Using Image Alignment and Disparity Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional autonomous vehicle systems face challenges in accurately detecting road debris using deep neural networks due to the scarcity of training data and the difficulty in encountering various types of road objects, limiting their ability to recognize arbitrary objects not predefined in their training lists.
Innovation Solution
A geometric approach for object detection using image alignment, which includes region of interest initialization, image alignment with homography estimation, and hazard detection and tracking, allowing for the detection of objects above a threshold size without prior knowledge of specific object types, by extracting sparse features and tracking them across frames.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used for object detection, then detection accuracy for predefined objects improves, but the system cannot detect arbitrary objects not in the predefined list
Solution Approach 1:
The system performs multiple functions: it detects both predefined objects (cars, pedestrians, cyclists) and arbitrary objects (road debris, cones, ladders) using a unified geometric approach. The homography-based detection method is universal and can identify any object on the road surface regardless of whether it was pre-defined in training data, making the system multi-functional without requiring separate detection mechanisms for different object types.
2Reliability
If real-world data is collected to train DNNs, then training data quality improves, but the scarcity of road objects in real-world environments limits data quantity
Solution Approach 1:
Instead of relying on scarce real-world captured images of road debris, the system uses synthetic disparity images generated through geometric processing of normal road images. The homography transformation creates virtual representations of road objects and disparities that simulate various debris scenarios, providing abundant training data without requiring physical collection of rare real-world instances.
3Loss of time
If conventional DNNs are trained on limited debris types, then training time reduces, but detection efficacy for other debris types decreases
Solution Approach 1:
The system changes the fundamental parameter of object detection from learning object-specific features through lengthy training to using geometric parameters (homography matrices, disparity values) that are universally applicable to all road objects. This parameter transformation eliminates the need for extensive training on specific debris types while maintaining high detection efficacy across diverse objects, as the geometric relationships remain consistent regardless of object identity.
4Adaptability or versatility
If physical vehicles are used to collect training data, then real-world scenario coverage improves, but safety risks and time consumption increase
Solution Approach 1:
The system creates virtual copies of real-world driving scenarios through image processing and homography transformations rather than physically driving vehicles to collect data. By processing existing road images through geometric transformations, the system generates synthetic training data that covers various scenarios without the time-consuming and safety-risking process of physical data collection campaigns.
Data Source
AI summary
A geometric approach may be used to detect objects on a road surface. A set of points within a region of interest between a first frame and a second frame are captured and tracked to determine a difference in location between the set of points in two frames. The first frame may be aligned with the second frame and the first pixel values of the first frame may be compared with the second pixel values of the second frame to generate a disparity image including third pixels. One or more subsets of the third pixels that have a value above a first threshold may be combined, and the third pixels may be scored and associated with disparity values for each pixel of the one or more subsets of the third pixels. A bounding shape may be generated based on the scoring.