Agent Trajectory Derivation via Ground Plane Tracking Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for collecting prior agent trajectories, such as those using LiDAR-based sensor systems, are expensive and limited in scalability due to the high cost of equipment and restricted geographic availability, making it impractical to collect accurate and large-scale data for applications like autonomous driving and transportation matching platforms.
Innovation Solution
The use of monocular cameras and stereo camera systems, integrated with telematics sensors, to capture and process image data for deriving agent trajectories, which involves identifying tracking points, estimating depth, and applying motion models to correct for occlusions and inaccuracies, allowing for more affordable and widespread data collection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LiDAR-based sensor systems are used to collect agent trajectories, then measurement precision is improved, but device complexity and cost increase significantly
Solution Approach 1:
The patent uses monocular camera images to create a simplified copy of the depth information that would otherwise require complex LiDAR systems. By identifying tracking points in 2D images and estimating their depth through geometric relationships with the ground plane, the system replicates the trajectory measurement function of LiDAR using much simpler, cheaper hardware.
Solution Approach 2:
The patent replaces the mechanical LiDAR scanning system with an optical camera-based approach. Instead of using rotating mirrors and laser beams to actively scan the environment, the system passively captures images with a monocular camera and derives trajectory information through image processing and geometric estimation algorithms.
2Measurement precision
If LiDAR-based sensor systems are deployed, then measurement precision is improved, but scalability is reduced due to high cost
Solution Approach 1:
The patent employs inexpensive monocular cameras that can be widely distributed across many vehicles rather than expensive LiDAR systems. These low-cost sensors can be deployed at scale across large fleets, enabling massive data collection efforts that would be economically infeasible with high-cost hardware, even if individual sensor performance is lower.
3Device complexity
If monocular cameras are used instead of LiDAR, then device complexity is reduced, but measurement precision deteriorates due to lack of direct depth information
Solution Approach 1:
The patent introduces the ground plane as an intermediary reference that mediates between the 2D image data and 3D spatial relationships. By establishing tracking points on the ground plane and using the known geometry of the ground, the system creates a reliable reference frame that enables accurate depth estimation from monocular images without requiring complex active sensing.
Solution Approach 2:
The patent transitions from 2D image coordinates to 3D world coordinates by introducing depth estimation through ground plane geometry. By mapping tracking points from the 2D image plane to 3D space using the ground plane as a reference, the system effectively adds the depth dimension to monocular vision data, enabling accurate trajectory measurement despite the inherent limitations of single-camera imaging.
4Duration of action of stationary object
If tracking points are identified in occluded images, then continuity of trajectory data is improved, but measurement precision may deteriorate due to occlusion errors
Solution Approach 1:
The patent prepares for potential occlusions by establishing robust tracking point identification methods that can handle partial visibility. By using pixel masks to precisely identify visible portions of agents and selecting tracking points based on ground plane intersections rather than arbitrary features, the system creates a cushion against occlusion errors that maintains tracking continuity even when agents are partially hidden.
Data Source
AI summary
Examples disclosed herein may involve a computing system that is operable to (i) receive a sequence of images captured by a camera associated with a vehicle, (ii) for each of at least a subset of the received images in which a given agent is detected, (a) generate a respective pixel mask that identifies a boundary of the given agent within the image, (b) identify, as a tracking point for the given agent within the image, at least one given pixel within the pixel mask that is representative of an estimated intersection point between the given agent and a ground plane, and (c) determine a position of the given agent at the capture time of the image based on the tracking point and information regarding the ground plane, and (iii) determine a trajectory for the given agent based on the determined positions of the given agent.