Multitask Object Detection for Occluded Object-Part Association
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks struggle with accurately identifying objects that are partially occluded in images, leading to incorrect detection of object parts, particularly in complex scenes like autonomous vehicle navigation.
Innovation Solution
A multitask object detection system that generates a feature map to identify objects and their parts, using a neural network to associate body parts of a person or other objects, even when partially occluded, by constructing a cost function and employing a Hungarian algorithm for accurate object association.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a convolutional neural network is used for object detection, then the system can identify objects and extract high-level features, but the detection accuracy deteriorates when objects are occluded or partially visible
Solution Approach 1:
The patent divides object detection into multiple tasks: detecting complete objects, detecting occluded objects, and detecting body parts. By segmenting the detection problem into these distinct tasks with separate neural network branches, the system can specialize each branch for specific detection scenarios, thereby improving overall accuracy for occluded objects and body parts
Solution Approach 2:
The patent adds a temporal dimension to object detection by using video input instead of static images. By analyzing sequences of frames and tracking object appearances over time, the system can infer occluded objects and body parts that are not visible in individual frames, improving detection reliability under occlusion conditions
2Adaptability or versatility
If the system attempts to detect all objects in a scene, then comprehensive object identification is achieved, but body part detection accuracy deteriorates due to occlusion and complexity
Solution Approach 1:
The patent implements task segmentation with separate detection branches for complete objects, occluded objects, and body parts. Each branch is optimized for its specific detection target, allowing the system to maintain high body part detection accuracy while still providing comprehensive object detection coverage through the combined output of all branches
Solution Approach 2:
The patent introduces an object association module that acts as an intermediary to correlate detection results from multiple branches and frames. This module uses cost functions and the Hungarian algorithm to match detected body parts with their corresponding objects, resolving ambiguities and improving body part detection accuracy in complex scenes
3Productivity
If traditional object detection methods are used, then processing speed is maintained, but the system fails to accurately associate body parts with their corresponding objects
Solution Approach 1:
The patent performs preliminary detection of objects, occluded objects, and body parts in parallel using separate neural network branches before performing association. By pre-processing and detecting all potential targets simultaneously, the system maintains high processing speed while providing complete information for subsequent association tasks
Solution Approach 2:
The object association module serves as an intermediary that efficiently matches body parts with objects using cost functions and the Hungarian algorithm. This dedicated association mechanism recovers lost object-part relationships without significantly impacting processing speed, as it operates on already-detected features rather than raw image data
Data Source
AI summary
System and techniques are described herein for processing images to detect objects in the provided images. In one illustrative example, a method of processing image data includes obtaining an image including at least a first object. The method can include generating a feature map based on providing the image to a neural network. The method can further include identifying a plurality of objects based on the feature map, the plurality of objects including a first part of the first object. The method can include identifying a first set of object parts within the plurality of objects corresponding to the first object.