Dynamic Object Detection Using LIDAR Tensor Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle systems face challenges in accurately detecting dynamic objects, particularly those not previously trained on, due to the sparse and non-uniform nature of LIDAR point clouds and missing appearance information, which can lead to incorrect corrective actions.
Innovation Solution
A computer-implemented method using a machine learning algorithm that processes an unlimited sequence of LIDAR point clouds, generating feature vectors and tensors to detect dynamic objects class-agnostically, incorporating odometry information to compensate for ego-motion, and predicting future locations and velocities for real-time control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If LIDAR point clouds are used for object detection, then detection range and 3D spatial awareness are improved, but the sparse and non-uniform nature of the data reduces measurement precision
Solution Approach 1:
The patent combines multiple LIDAR point clouds acquired at different time points to form a composite representation. By merging sequential point cloud data, the system accumulates sufficient object information despite the sparse nature of individual scans, thereby improving detection precision while maintaining extended detection range.
Solution Approach 2:
The system performs preliminary processing of LIDAR point clouds by organizing and storing them in advance before object detection is needed. Point clouds are pre-processed, aligned, and stored in a buffer, enabling the detection algorithm to work with prepared data structures rather than raw sequential scans, thus improving measurement precision.
2Measurement precision
If machine learning algorithms are trained on specific object classes, then detection accuracy for known objects is improved, but the system cannot detect objects not previously trained on
Solution Approach 1:
The patent implements a universal object detection framework that can handle both known and unknown object classes. The system uses unsupervised learning techniques and anomaly detection algorithms that do not require pre-training on specific object classes, enabling the same system to detect both familiar and novel objects with comparable effectiveness.
Solution Approach 2:
The system dynamically adjusts detection parameters and thresholds based on the characteristics of detected objects. When an object does not match known classes, the system modifies its detection parameters to accommodate novel object characteristics, thereby maintaining detection accuracy while improving adaptability to unknown objects.
3Reliability
If the system processes unlimited sequences of LIDAR point clouds, then detection completeness is improved, but processing time and computational resources increase
Solution Approach 1:
The system processes a partial sequence of LIDAR point clouds rather than unlimited sequences. By limiting the processing window to a predetermined number of recent point clouds, the system achieves sufficient detection completeness for safety-critical applications while constraining computational resources and processing time to acceptable levels.
Solution Approach 2:
The system implements real-time processing by skipping older point cloud data that is no longer relevant to current detection needs. Point clouds beyond a certain time threshold are discarded without detailed processing, allowing the system to maintain high detection completeness for current objects while reducing overall processing time and computational load.
Data Source
AI summary
A method for determining a set of dynamic objects in sensor data representative of a surrounding area of a vehicle having sensors, the method being executed by a server, the server executing a machine learning algorithm (MLA). Sensor data is received, and the MLA generates, based on the sensor data, a set of feature vectors. Vehicle data indicative of a localization of the vehicle is received. The MLA generates, based on the set of feature vectors and the vehicle data, a tensor, the tensor including a grid representation of the surrounding area. The MLA generates an mobility mask indicative of grid cells occupied by at least one moving potential object in the grid, and a velocity mask indicative of a velocity associated with the at least one potential object in the grid. The MLA determines, based on the mobility mask and the velocity mask, the set of dynamic objects.