Pedestrian Position Prediction Using 2D Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current pedestrian detection systems in ADAS vehicles rely on costly and complex sensors like stereo cameras, LIDARs, and depth sensors, making them unsuitable for lower-priced vehicles, and lack methods for predicting the future position of pedestrians using only 2D camera vision.
Innovation Solution
A method that captures 2D video feed, detects objects, obtains their pose using key points, determines movement characteristics, and predicts future positions using a Kalman filter, without requiring 3D information or GPS, allowing for cost-effective and simple hardware configurations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If stereo cameras, LIDARs, radars, and depth sensors are used for pedestrian detection, then measurement precision and reliability are improved, but device complexity and cost increase
Solution Approach 1:
The patent extracts and removes the complex 3D sensing components (stereo cameras, LIDARs, radars, depth sensors) from the system, achieving pedestrian detection and future position prediction using only a simple 2D vision camera. This extraction eliminates the need for expensive multi-sensor configurations while maintaining detection capability through 2D image processing and pose estimation algorithms
Solution Approach 2:
The patent creates a 2D pose representation (skeleton model with key points) that serves as a simplified copy of the pedestrian's spatial configuration. Instead of using complex 3D depth information, the system captures essential movement characteristics through 2D key point coordinates, enabling accurate prediction without expensive sensors
2Measurement precision
If 3D camera arrangements, LIDARs, and depth sensors are combined for position prediction, then prediction accuracy is improved, but ease of manufacture deteriorates
Solution Approach 1:
The patent removes complex 3D sensing hardware (LIDARs, depth sensors, GPS) from the system architecture, achieving future position prediction using only a standard 2D vision camera. This extraction simplifies manufacturing and deployment while maintaining prediction accuracy through sophisticated 2D pose-based algorithms
Solution Approach 2:
The patent replaces expensive, complex sensing systems with inexpensive 2D camera hardware that is widely available and easy to manufacture. The system uses computationally intensive but hardware-simple processing to achieve predictions that previously required expensive sensor fusion
3Reliability
If ADAS systems include sophisticated sensors for pedestrian detection, then safety and reliability are improved, but device complexity increases
Solution Approach 1:
The patent extracts the essential function of pedestrian detection and prediction from complex multi-sensor ADAS systems, achieving reliable collision avoidance using only a 2D vision camera. This extraction maintains safety-critical functionality while eliminating unnecessary sensor complexity
Solution Approach 2:
The patent substitutes complex hardware-based 3D sensing mechanisms with software-based 2D image processing and pose estimation algorithms. Instead of using multiple physical sensors to capture depth information, the system uses computational methods to infer 3D pose and predict future position from 2D images
Data Source
Figure 1~3
Figure 4~5
Figure 6
AI summary
The method for predicting a future position of an object comprises capturing a video feed including a two-dimensional (2D) image data; detecting an object in the two-dimensional (2D) image data; obtaining a pose of the object which comprises a plurality of key points; generating a history of poses of the object; determining movement characteristics of two or more key points of the object; obtaining a predicted position data of each of the two or more key points by at least estimating an estimated state of each of the two or more key points based on the movement characteristics and the history of poses; and predicting a future position of at least a predefined representative point of the object from the predicted position data, wherein said future position is predicted on a two-dimensional (2D) image data of the video feed.