SLAM Object Tracking With Sensor Fusion for Uncertain Detections
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional SLAM methods face challenges in handling uncertainties and noisy object detections, particularly in determining and tracking objects in environments using mobile devices, as they rely solely on geometrical information and struggle with ambiguous data allocation.
Innovation Solution
The method employs sensor data from lidar, cameras, and inertial sensors to detect objects using deep neural networks, synchronizes and preprocesses data, and implements object tracking by transforming detected objects into a shared reference coordinate system, clustering multiple detections into real objects, and updating SLAM graphs with uncertainty estimation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional SLAM methods use only geometrical information for object detection, then the system complexity is low, but the reliability of object allocation deteriorates due to noisy detections and uncertainties
Solution Approach 1:
The patent segments object detection into two distinct stages: geometrical detection using SLAM for position and shape, and semantic detection using deep neural networks for classification. This segmentation allows each detector to specialize, improving overall reliability while managing complexity through modular architecture.
Solution Approach 2:
The patent merges geometrical SLAM detection results with semantic deep learning detection results by combining their outputs through a unified object representation that includes both geometric parameters and semantic class labels, thereby improving allocation reliability through complementary information.
2Reliability
If deep neural networks are used for semantic object detection, then the reliability of object classification improves, but the processing time and computational complexity increase
Solution Approach 1:
The patent performs geometrical detection and obtains basic object parameters (position, orientation, dimensions) through SLAM before applying semantic classification. This preliminary geometrical filtering reduces the search space and allows semantic networks to focus only on relevant detections, reducing overall processing time.
Solution Approach 2:
The patent applies semantic detection selectively rather than to all sensor data - only to geometrical detections that meet certain confidence thresholds or are deemed relevant based on the environment and task, thereby reducing computational overhead while maintaining classification reliability for critical objects.
3Measurement precision
If multiple sensors (lidar, cameras, inertial sensors) are integrated for data acquisition, then the measurement precision of environment and object detection improves, but the device complexity increases
Solution Approach 1:
The patent implements a unified sensor fusion framework that processes data from multiple sensor types (lidar, cameras, inertial sensors) through a common SLAM and object detection pipeline. This universal processing architecture handles diverse sensor inputs consistently, improving measurement precision while managing integration complexity through standardized interfaces.
Solution Approach 2:
The patent introduces a coordinate transformation and synchronization module as an intermediary between raw sensor data and the object detection pipeline. This intermediary standardizes different sensor data formats, synchronizes timing, and transforms coordinates, thereby enabling precise multi-sensor integration without proportionally increasing overall system complexity.
4Reliability
If objects are tracked across multiple SLAM datasets with coordinate transformations, then the robustness of object allocation improves, but the computational complexity increases
Solution Approach 1:
The patent performs coordinate transformations and transforms object detections into a common reference frame as a preliminary step before object allocation and tracking. By pre-transforming all detections to a unified coordinate system, the subsequent allocation algorithm operates on standardized data, improving robustness while reducing the computational complexity of the main tracking loop.
Data Source
AI summary
A method for determining objects in an environment with the aid of SLAM and a mobile device in the environment, which has at least one sensor for acquiring object and/or environment information. The method includes: providing sensor data, carrying out an object detection in order to obtain first object datasets for detected objects; carrying out object tracking for a new SLAM dataset, including allocating objects detected with the aid of the object detection to real objects in order to obtain second object datasets or real objects to be considered in the SLAM graph.