Self-Driving Car Object Detection Using Local-Global Point Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing self-driving car systems face challenges in accurately and quickly detecting objects in their surroundings, particularly in scenarios where false negatives occur, which can lead to potential collisions.
Innovation Solution
The system employs a combination of neural networks to generate feature vectors that incorporate both local and global contexts from LiDAR point clouds, using Convolutional Neural Networks (CNNs) to enhance object detection by integrating local geometry and surrounding environmental information, allowing for improved trajectory planning and collision avoidance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional sensor systems and detection algorithms are used, then the system structure remains simple, but the object detection accuracy and speed are insufficient, leading to false negatives
Solution Approach 1:
The patent combines LIDAR point cloud data with camera image data to create a fused representation of the environment. This multi-sensor fusion approach integrates different data types (3D spatial information from LIDAR and 2D visual information from camera) to improve detection accuracy while managing system complexity through unified processing architecture
Solution Approach 2:
The patent transforms 2D image data into 3D spatial representations by projecting image features onto the LIDAR point cloud framework. This dimensional transformation allows the system to leverage both 2D visual recognition capabilities and 3D spatial understanding, enhancing object detection accuracy through multi-dimensional feature integration
2Speed
If traditional detection algorithms are used, then the processing speed is sufficient for basic operations, but the detection time is too long to respond to fast-moving objects like running children
Solution Approach 1:
The patent pre-processes LIDAR point cloud data to generate clustered point representations and pre-computes spatial features before object detection is needed. This preliminary organization of data structures enables faster query and processing during actual detection events, reducing latency for time-critical scenarios
Solution Approach 2:
The patent divides the point cloud data into clustered segments representing different objects or regions of interest. This segmentation allows the detection algorithm to process discrete object groups independently rather than analyzing the entire point cloud, significantly improving processing speed while maintaining detection accuracy through focused analysis
3Loss of information
If simple binary detection is used, then the system operation remains simple, but the system cannot provide sufficient information for trajectory planning and safe vehicle operation
Solution Approach 1:
The patent generates bounding box annotations with varying levels of detail and precision tailored to different objects and their importance. Critical objects receive more detailed spatial characterization and feature extraction, while less important objects receive simplified processing. This differential information processing reduces overall complexity while ensuring sufficient information is available for safety-critical decisions
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method (1000) and electronic device (210) for detecting an object are disclosed. The method (1000) includes generating a cluster of points representative of the surroundings of the SDC (220), generating by a first Neural Network (NN) a first feature vector based on the cluster indicative of a local context of the given object in the surroundings of the SDC (220), generating by a second NN second feature vectors for respective points from the cluster based on a portion of the point cloud, where a given second feature vector is indicative of the local and global context of the given object, generating by the first NN a third feature vector for the given object based on the second feature vectors representative of the given object, and generating by a third NN a bounding box around the given object using the third feature vector.