UAV Object Detection via Split Image Processing and Box Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using convolution neural networks in UAV images are limited by their complexity, leading to reduced accuracy and real-time performance, especially at higher altitudes and when detecting small objects.
Innovation Solution
The proposed method employs a lightweight convolution neural network with split image processing (SIP) and box merging techniques to improve object detection accuracy and speed by splitting input images into overlapping left and right images, resizing, normalizing, and merging bounding boxes based on coordinate values and object size criteria.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a complex convolution neural network model is used to improve object detection accuracy in UAV images, then detection accuracy improves, but real-time performance deteriorates due to high computational requirements
Solution Approach 1:
The patent applies segmentation by dividing the image processing task into multiple stages: preprocessing (resize, normalize), feature extraction through multiple convolutional layers with increasing depth, and post-processing (non-maximum suppression). This staged approach allows the system to progressively build detection accuracy while managing computational load at each stage, enabling real-time performance.
Solution Approach 2:
The patent introduces multi-scale feature detection by processing images at different resolutions and scales. The object detector analyzes features at multiple depth levels and scales, allowing it to detect objects of varying sizes efficiently. This dimensional approach to feature extraction improves accuracy without requiring a single overly complex model.
2Measurement precision
If a complex convolution neural network model is used to improve object detection accuracy in UAV images, then detection accuracy improves, but device complexity increases
Solution Approach 1:
The object detector is segmented into distinct functional components: multiple convolutional layers for feature extraction, pooling layers for downsampling, and a detection head for final object classification. This modular segmentation allows each component to be optimized independently, reducing overall system complexity while maintaining high detection accuracy.
Solution Approach 2:
The patent employs dynamic feature extraction where the network adaptively adjusts which features to extract at different depths and scales based on the input image characteristics. This dynamic approach allows the model to focus computational resources on relevant features, reducing unnecessary complexity in the model structure.
3Measurement precision
If existing object detection methods are used in UAV images at higher altitudes, then comprehensive object coverage is achieved, but detection accuracy deteriorates due to small object size
Solution Approach 1:
The patent addresses small object detection by introducing multi-scale analysis. The object detector processes features at multiple scales and depths, allowing it to effectively detect and classify small objects that would be invisible at single scale. This multi-dimensional feature space expansion enables accurate detection of small objects captured from high altitudes.
Solution Approach 2:
The patent applies local quality enhancement by using different convolution kernel sizes and pooling operations at different stages of the network. Early layers use smaller kernels to preserve fine details of small objects, while deeper layers use larger receptive fields to capture contextual information. This localized optimization of feature extraction improves small object detection accuracy.
Data Source
AI summary
According to the present disclosure, an apparatus for real-time object detection in an image captured by an unmanned aerial vehicle (UAV) including an image splitter configured to receive an input image and split the input image into a left image and a right image; and an object detector configured to detect an object from each of the left image and the right image provided from the image splitter is provided.