Rear-Surface Vehicle Detection for Accurate Adjacent-Lane Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous driving systems face challenges in accurately calculating the distance to vehicles in neighboring lanes due to the inclusion of side surfaces in the detected vehicle area, leading to inaccurate width measurements.
Innovation Solution
An electronic device for autonomous vehicles uses a camera, processor, and vehicle detection model to generate a bounding box, sliding window, and extended bounding box, calculating pixel differences to identify the center of the rear surface of a target vehicle, thereby improving distance calculation accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the vehicle detection model detects the vehicle area including side surfaces, then the vehicle detection coverage is improved, but the width measurement accuracy deteriorates
Solution Approach 1:
The detected vehicle area is segmented into multiple regions: a first region corresponding to the rear surface of the target vehicle and a second region corresponding to the side surface. By dividing the bounding box area into these distinct segments, the system can selectively process only the relevant rear surface region for width measurement, thereby maintaining detection coverage while improving measurement accuracy.
Solution Approach 2:
The method extracts the rear surface region from the complete vehicle bounding box by identifying characteristic features (such as the horizontal line corresponding to the rear bumper and vertical lines corresponding to rear pillars). This extraction isolates the measurement-critical rear surface area, removing the side surface portions that would otherwise contaminate the width measurement.
2Area of stationary object
If the bounding box includes the side surface of the target vehicle, then the complete vehicle area is captured, but the distance calculation accuracy deteriorates
Solution Approach 1:
The bounding box area is segmented into a first region (rear surface) and a second region (side surface). The width calculation is performed exclusively on the first region by measuring the horizontal distance between the left and right edges of the rear surface, excluding the side surface portions in the second region. This segmentation enables complete area capture while ensuring accurate distance calculation.
Solution Approach 2:
Different regions of the bounding box are assigned different functional qualities: the first region (rear surface) is designated for precise width measurement, while the second region (side surface) is excluded from measurement calculations. This local differentiation ensures that the measurement process focuses only on the geometrically accurate rear surface area.
3Measurement precision
If the sliding window width is reduced to half of the bounding box width, then the rear surface detection precision is improved, but the processing complexity increases
Solution Approach 1:
The sliding window approach divides the bounding box into multiple processing zones by using a window width equal to half the bounding box width. This segmentation allows the algorithm to systematically scan and analyze the rear surface region in discrete steps, improving detection precision through localized analysis while maintaining manageable processing complexity through structured iteration.
Data Source
AI summary
An electronic device provided in an autonomous vehicle, the electronic device comprising a camera, a memory storing at least one instruction, and at least one processor operatively coupled with the camera, wherein the at least one processor is configured to, when the at least one instruction is executed obtain a front image in which the autonomous vehicle is driving through the camera, identify a target vehicle in the front image based on the vehicle detection model stored in the memory, generate a bounding box corresponding to the target vehicle in response to an identification of the target vehicle, generate a sliding window having a height equal to the height of the bounding box and having a width half of the width of the bounding box, divide the bounding box into a first area positioned left based on a middle position of the width of the sliding window and a second area positioned right based on the middle position, generate an extended bounding box by extending the first area in a left direction and extending the second area in a right direction, wherein size of the extended bounding box is twice as wide as size of the bounding box, obtain a sum of a pixel difference values between the first area and the second area for each shift by sequentially shifting the sliding window by a predefined pixel interval with respect to all of width of the extended bounding box, and identify a point that corresponds to a minimum value among sum values respectively indicating the sums that are obtained according to the shifting.