Vehicle Object Detection Using Multi-Scale Feature Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving systems are computationally expensive and lack robust data for vehicle control, necessitating improved imaging configurations and devices that can efficiently detect objects and determine their distance without relying on high-cost assistive data devices like LIDAR.
Innovation Solution
The system employs an image sensor and an object detection module with a processor configured to receive image data, detect objects, determine their category, localize regions of interest, and calculate distance values based on trained parameters and multi-scale feature encoding layers, allowing for accurate object detection and distance determination without the need for LIDAR, and outputs these values for vehicle control systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pixel-level mapping is used to detect objects, then object detection capability is improved, but computational cost increases significantly
Solution Approach 1:
The patent segments the object detection process into distinct stages: region proposal generation, classification, and distance estimation. By dividing the image into regions of interest first, then processing only those regions, the system reduces overall computational load while maintaining detection accuracy
Solution Approach 2:
The patent extracts only the necessary features for object detection and distance estimation, rather than processing all pixel-level data. By focusing on key regions and their characteristics, the system achieves accurate detection with reduced computational requirements
2Device complexity
If conventional imaging systems are used without assistive devices, then device complexity is reduced, but measurement precision of distance deteriorates
Solution Approach 1:
The patent introduces an intermediary computational process that uses multi-scale feature encoding and object category information to infer distance. This intermediary layer processes image features and combines them with object size knowledge to estimate distance, achieving accurate measurements without direct distance sensors
Solution Approach 2:
The patent changes the approach from direct distance measurement to indirect distance estimation through multiple parameters. By analyzing object category, detected size, and position in the image, the system infers distance through parameter relationships rather than direct measurement
3Measurement precision
If more assistive data devices are added, then measurement precision is improved, but device complexity and cost increase
Solution Approach 1:
The patent makes the image sensor perform multiple functions: object detection, classification, localization, and distance estimation. By enabling the single imaging device to execute multiple tasks through advanced processing, the system eliminates the need for separate distance measurement devices
Solution Approach 2:
The system uses its own image data and processed features to determine distance, rather than relying on external assistive devices. The image processing pipeline extracts distance information from the same image data used for object detection, making the system self-sufficient
Data Source
AI summary
The present disclosure relates to object detection and object distance determination by an assisted driving system of a vehicle. In one embodiment, a method includes receiving image data detected and detecting at least one object in the image data. Detecting includes determining a region of interest in the image, classifying the object and localizing the region of interest in the image data. The method may include determining a distance value for the at least one object, wherein the distance determination is determined based on the category determined, localization of the region of interest, and object size relative to the image data. The method may include outputting a distance value determination for the at least one object.