Multi-Channel Object Detection Network for Occluded Vehicle Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in making accurate object predictions from raw sensor data without relying on human-programmed logic, particularly in occluded scenarios and adverse weather conditions, where traditional approaches may fail due to limited sensor visibility and data quality.
Innovation Solution
A fully-learned neural network system that processes multiple channels of input sensor data, including height, intensity, second return, and elongation channels, to predict object occupancy in space, using co-training techniques with human-labeled and auto-labeled data, and systematically altering training data to enhance robustness and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional human-programmed logic is used to process sensor data, then the system is easier to interpret and control, but the accuracy of object predictions in occluded scenarios and adverse weather conditions deteriorates
Solution Approach 1:
The patent replaces traditional human-programmed logic (mechanical/systematic approach) with a neural network system that learns patterns from data. The neural network substitutes rule-based processing with learned representations, enabling accurate object predictions in occluded scenarios and adverse weather where traditional logic fails.
Solution Approach 2:
The patent transforms sensor data into multiple channels (height, intensity, second return, elongation) and processes them through a neural network with multiple layers. This parameter transformation and multi-layer processing enables the system to extract meaningful features from raw sensor data, improving prediction accuracy despite increased computational complexity.
2Reliability
If multiple channels of sensor data are processed, then the robustness to sensor failures and adverse weather improves, but the computational complexity and data processing requirements increase
Solution Approach 1:
The patent combines multiple channels of sensor data (height, intensity, second return, elongation) into a unified neural network processing pipeline. By merging these diverse data sources, the system achieves robustness to sensor failures and adverse weather conditions, as the neural network can learn complementary patterns across channels.
Solution Approach 2:
The neural network is designed to process multiple types of sensor data simultaneously through a universal architecture. The same network structure handles different data channels and various operating conditions (occluded scenarios, adverse weather), making the system multi-functional and robust without requiring separate processing pipelines for each condition.
3Measurement precision
If co-training techniques with human-labeled and auto-labeled data are used, then the accuracy and robustness of the neural network improves, but the training time and computational resources increase
Solution Approach 1:
The patent employs co-training techniques where the neural network is pre-trained on large volumes of auto-labeled data to learn general patterns, then fine-tuned on smaller sets of human-labeled data to achieve high accuracy. This preliminary action with auto-labeled data reduces the time required for final training compared to using only human-labeled data from scratch.
Solution Approach 2:
The co-training approach maintains continuous learning by alternating between training on auto-labeled data and human-labeled data. This continuous useful action allows the network to progressively improve accuracy while efficiently utilizing available computational resources, balancing training time and performance gains.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating object predictions using a neural network. One of the methods includes receiving respective projections of a plurality of channels of input sensor data, wherein each channel of input sensor data represents different respective characteristics of electromagnetic radiation reflected off of one or more objects. Each of the projections of the plurality of channels of input sensor data are provided to a neural network subsystem trained to receive projections of input sensor data as input and to provide an object prediction as an output. At the output of the neural network subsystem, an object prediction that predicts a region of space that is likely to be occupied by an object is received.