Ego Vehicle Speed Estimation via 3D-CMA Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for ego vehicle speed estimation rely heavily on in-car sensors and require camera calibration, limiting their effectiveness in real-world applications, especially in scenarios where fixed views are not feasible.
Innovation Solution
A 3D Convolutional Neural Network (3D-CNN) with masked-attention (3D-CMA) architecture processes video frames from a monocular camera to estimate ego vehicle speed by focusing on relevant road features, such as lane line segments, using a masked-attention mechanism to filter out irrelevant motion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If in-car sensors and CAN bus are used for speed extraction, then speed estimation reliability is improved, but device complexity and dependency on additional hardware increase
Solution Approach 1:
The patent replaces the mechanical/sensor-based speed measurement system with a vision-based computational system. Instead of using in-car sensors and CAN bus hardware, the invention uses a monocular camera combined with a 3D-CMA neural network to estimate speed by analyzing optical flow and road feature displacement, thereby eliminating hardware dependency while maintaining reliability
Solution Approach 2:
The patent introduces masked-attention maps as an intermediary mechanism that guides the neural network to focus on relevant road features (such as lane markings) while filtering out irrelevant motion. This intermediary attention mechanism enables accurate speed estimation from visual data alone, replacing the need for direct sensor input
2Measurement precision
If camera calibration and fixed view are required, then measurement precision is improved, but adaptability to different scenarios deteriorates
Solution Approach 1:
The patent employs a dynamic masked-attention mechanism that adapts to different driving scenarios in real-time. The attention masks are generated dynamically based on the current video input, allowing the system to automatically focus on relevant road features regardless of the specific scenario, eliminating the need for fixed camera views or pre-calibration
Solution Approach 2:
The 3D-CMA architecture is designed to be universally applicable across different driving conditions and scenarios. The masked-attention mechanism enables the same system to accurately estimate speed in various environments (urban, highway, rural) without requiring scenario-specific calibration or multiple fixed cameras, achieving multi-functionality from a single monocular camera
Data Source
AI summary
A system, apparatus and method of embedded ego vehicle speed estimation apparatus, including a car-mounted monocular camera for capturing a sequence of video frames of an outdoor scene from a moving car, where the outdoor scene includes a road, as a camera channel, and processing circuitry. The processing circuitry is configured with a single-shot network and a 3D convolutional neural network (3D-CNN), the single-shot network segments features of the road in the video frame sequence and generates a masked-attention map for the segmented road features, a concatenation circuit concatenates the masked-attention map as an additional channel to the camera channel to generate a masked-attention input, and the 3D-CNN network receives the masked-attention input and generates an estimated speed of the ego vehicle based on displacement of the segmented road features in the video sequences.