Autonomous Vehicle Multi-Object Tracking With PMBM Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current multi-object tracking systems in autonomous vehicles face challenges in effectively handling uncertainties regarding the number of objects, their locations, and the 'birth and death' phenomena, particularly with lidar and 2D/3D detections, leading to inefficiencies in data processing and false positives.
Innovation Solution
The implementation of Poisson multi-Bernoulli mixture filtering to reduce data processing by associating object identifiers with probabilities, using a system that includes a camera, lidar, and a bounding box determination module to store and compare object data, forming hypotheses and generating probability data to control vehicle operations, thereby reducing the number of objects and hypotheses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Poisson multi-Bernoulli mixture filtering is used to model unknown number of objects and birth-death phenomena, then the ability to handle uncertainties in object tracking is improved, but the computational complexity increases
Solution Approach 1:
The patent segments the complex Poisson multi-Bernoulli mixture filtering problem into distinct components: detected objects, undetected objects, and hypotheses. Each component is processed separately through probability generation and filtering steps, making the overall complex system manageable and computationally tractable while maintaining reliability in handling uncertainties
Solution Approach 2:
The patent applies partial action by generating probability data for all objects and hypotheses but then selectively filtering based on probability thresholds. This approach computes comprehensive probability information but only retains and processes the most relevant results, balancing thorough uncertainty modeling with computational efficiency
2Measurement precision
If comprehensive probability data is generated for all detected and undetected objects, then the accuracy of object tracking is improved, but the data processing time increases
Solution Approach 1:
The patent performs preliminary probability generation for all objects and hypotheses before final filtering and decision-making. By pre-computing probability data and organizing it into structured sets, the system prepares comprehensive accuracy information in advance, enabling faster real-time tracking decisions without sacrificing precision
Solution Approach 2:
The patent changes the state of raw object data by associating probability parameters with each object and hypothesis. This transformation converts uncertain detection data into quantified probability measurements, enabling efficient filtering and comparison operations that maintain accuracy while reducing processing time through parameter-based decision rules
3Object-generated harmful factors
If hypotheses are reduced based on probability thresholds, then the number of false positives is decreased, but the risk of missing valid objects increases
Solution Approach 1:
The patent implements feedback through iterative hypothesis reduction based on probability thresholds. The system generates probability data, filters hypotheses below thresholds, and uses the filtered results to inform subsequent tracking decisions. This feedback loop continuously refines the set of tracked objects, reducing false positives while maintaining reliability through probability-based validation at each stage
Data Source
AI summary
A method and system for multi-object tracking is set forth. Object data for boundaries of a plurality of objects are received. Poisson multi-Bernoulli mixture filtering is performed on the object data to form a filtered set of object data. Ultimately, the filtered set of object data is used to control the operation of the vehicle. Identifiers and probabilities are associated with the objects to reduce the set of object data.