3D Object Detection Using Horizontal Regularization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current 3D object detection methods based on single frame images for ADAS have low accuracy and fail to adequately consider horizontal information, while methods using lidar or multi-frame image parallax are costly and power-intensive, unable to meet the requirements for low power consumption and cost-effective solutions.
Innovation Solution
A real-time object detection and 3D localization method that constrains the prediction of horizontal information using a regularization term in the deep neural network training process, incorporating spatial geometric correlation to enhance the accuracy and convergence speed of 3D object detection and localization, specifically by defining a weight matrix to maintain horizontal adjacency and adding a regularization term to the loss function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If single frame image 3D object detection is used, then power consumption and hardware cost are reduced, but detection accuracy is insufficient
Solution Approach 1:
The patent transforms the 2D image detection problem into a 3D space prediction problem. By predicting 3D bounding box parameters (position, size, orientation) directly from 2D image features, the method incorporates spatial geometric constraints that improve detection accuracy while maintaining single-frame processing efficiency and low power consumption requirements
Solution Approach 2:
The patent changes the prediction parameters from simple 2D bounding box coordinates to comprehensive 3D parameters including position (x, y, z), size (length, width, height), and orientation (roll, pitch, yaw). This parameter expansion enables the model to capture horizontal information and spatial geometric correlations, significantly improving detection accuracy without requiring multiple frames or expensive sensors
2Ease of manufacture
If existing single frame image algorithms are used, then hardware cost is low, but horizontal information estimation is inadequate
Solution Approach 1:
The patent addresses horizontal information loss by predicting 3D bounding box parameters that include horizontal position and orientation information. The model learns to infer horizontal spatial relationships from 2D image features, effectively recovering horizontal information that would otherwise be lost in single-frame processing
Solution Approach 2:
The patent incorporates loss functions that provide feedback on horizontal information accuracy during training. By optimizing for 3D bounding box prediction accuracy including horizontal parameters, the model receives continuous feedback to improve its estimation of horizontal information, achieving better performance without increased hardware cost
3Measurement precision
If multi-frame image parallax method is used, then 3D detection accuracy is improved, but sensor overhead and computational complexity increase
Solution Approach 1:
The patent extracts and utilizes horizontal information directly from single-frame 2D image features, eliminating the need for multi-frame parallax processing. By focusing on extracting relevant horizontal cues from the available single frame, the method achieves competitive accuracy without the sensor overhead and computational complexity of multi-frame approaches
Solution Approach 2:
The patent creates a 3D representation (virtual copy) of the object from 2D image data by predicting 3D bounding box parameters. This virtual 3D model captures horizontal information and spatial relationships without requiring actual multi-frame capture or complex parallax computation, achieving accuracy improvement with reduced device complexity
Data Source
AI summary
The invention relates to a real-time object detection and 3D localization method based on a single frame image. Comprising following steps: S1: inputting a 2D RGB image; S2: performing feature extraction on the 2D RGB image, extracting features of a deep network and a shallow network respectively; S3: carrying out 2D object detection and applying to subsequent modules; S4: estimating vertices, instance-level depth and center point of a 3D-box respectively; S5: adding a regularization term for maintaining horizontal locality into prediction of center point of a 3D-box to constrain and optimize the prediction of center point of the 3D-box; and S6: outputting a 2D RGB image with a 3D-box tag in combination with predictions of all modules. The invention increases the speed of model training convergence and the accuracy of 3D object detection and localization, and meets the accuracy requirements of an Advanced Driver Assistant System (ADAS) with a low hardware cost.