Pedestrian Motion Prediction Using Grid-Based Occupancy Contours
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in predicting the motion of pedestrians due to low-quality data and the complexity of their behavior, making it difficult to detect, classify, and track them effectively, especially in diverse and dynamic environments.
Innovation Solution
A grid-based prediction method is used to forecast the possible future locations of pedestrians, where a grid is projected around each detected pedestrian, with cell probabilities indicating the likelihood of movement based on observed speed, direction, and environmental factors, creating a heat map that aids in path planning and propagates perception uncertainty.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learned trajectory proposal based behavior models are used to predict pedestrian motion, then predictions can be made when data quality is high and feasible paths are limited, but the system fails when data quality is low or pedestrian behavior is complex and diverse
Solution Approach 1:
The prediction space is segmented into a grid of cells surrounding the pedestrian. Each cell is evaluated independently to determine the likelihood of the pedestrian entering it, allowing the system to handle complex and diverse pedestrian behaviors by breaking down the continuous prediction problem into discrete, manageable units.
Solution Approach 2:
The system changes the parameter representation from continuous trajectory proposals to discrete grid cell probabilities. This parameter transformation allows the system to effectively model uncertain and diverse pedestrian behaviors by representing predictions as a distribution of probabilities across multiple possible locations rather than a single deterministic path.
2Productivity
If traditional trajectory modeling is used, then predictions work for limited feasible paths, but the system cannot handle the infinite diversity and quick direction changes of pedestrians
Solution Approach 1:
The continuous space around the pedestrian is divided into a finite grid of cells. This segmentation transforms the infinite complexity of pedestrian motion into a manageable set of discrete possibilities, maintaining computational efficiency while capturing diverse behavior patterns.
Solution Approach 2:
The system adds a probability dimension to the prediction output, transforming single-point trajectory predictions into multi-point probability distributions across grid cells. This dimensional expansion allows the system to represent uncertainty and diversity in pedestrian behavior while maintaining computational tractability.
3Measurement precision
If high-quality data and ground truth training are required for accurate predictions, then prediction accuracy improves, but the system becomes less capable of handling low-quality real-world data
Solution Approach 1:
The system changes from requiring precise ground truth trajectories for training to learning from imprecise, real-world sensor data. By representing predictions as probability distributions over grid cells rather than precise continuous trajectories, the system can effectively learn from lower-quality data while maintaining useful prediction accuracy.
Solution Approach 2:
The system uses readily available, low-cost sensor data from the vehicle's perception system without requiring expensive, high-precision ground truth annotations. The grid-based approach allows effective learning from abundant but imperfect real-world data, eliminating the need for costly data collection and annotation processes.
Data Source
AI summary
The technology relates to controlling a vehicle (100) in an autonomous driving mode. For instance, sensor data identifying an object in an environment of the vehicle may be received. A grid (780, 782, 784) including a plurality of cells may be projected around the object. For each given one of the plurality of cells, a likelihood that the object will enter the given one within a period of time into the future is predicted. A contour (1280-1284) is generated based on the predicted likelihoods. The vehicle is then controlled in the autonomous driving mode in order to avoid an area within the contour.