Object Recognition Complexity Reduction via Action Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current human interacted object recognition systems are complex, leading to high processing and network transmission bandwidth requirements, making them inefficient and costly for real-time object detection in smart activity spaces.
Innovation Solution
The system reduces spatio-temporal complexity by using human action and pose contexts to select a subset of frames and regions with high object interaction probability, employing 3D cameras for skeletal joint tracking and Hidden Markov Models for action recognition, and refining object localization through change detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full-frame and full-video recognition is performed, then object recognition accuracy is improved, but processing complexity and network bandwidth requirements increase significantly
Solution Approach 1:
The patent segments the video processing task by dividing it into two stages: (1) action recognition at the frame level to identify temporal segments containing human-object interactions, and (2) object recognition only in those identified segments. This segmentation avoids processing the entire video stream, reducing computational complexity while maintaining recognition accuracy for relevant objects.
Solution Approach 2:
The patent performs preliminary action recognition on video frames to identify which frames contain human-object interactions before performing object recognition. This preliminary action filters the video stream to only those frames relevant for object detection, significantly reducing the amount of data that requires complex object recognition processing.
2Measurement precision
If full-frame and full-video recognition is performed, then object recognition accuracy is improved, but network transmission bandwidth requirements increase
Solution Approach 1:
The patent segments the video data transmission by sending only the identified temporal segments containing human-object interactions to the object recognition system, rather than transmitting the entire video stream. This segmentation reduces network bandwidth requirements while ensuring that all relevant object information is captured for accurate recognition.
Solution Approach 2:
The patent extracts and transmits only the relevant temporal segments from the video stream that contain human-object interactions, removing unnecessary frames and data. This extraction process reduces the quantity of data requiring network transmission while preserving all information necessary for accurate object recognition.
3Measurement precision
If action recognition and object recognition are performed jointly, then recognition accuracy is improved, but processing time increases
Solution Approach 1:
The patent segments the recognition process into two sequential stages: first performing action recognition to identify temporal segments with human-object interactions, then performing object recognition only in those segments. This segmentation eliminates redundant processing in non-relevant frames, reducing total processing time while maintaining recognition accuracy through the two-stage approach.
Solution Approach 2:
The patent performs preliminary action recognition to identify which temporal segments require object recognition, before performing the more computationally intensive object recognition task. This preliminary filtering action reduces the total processing time by avoiding object recognition in frames where no human-object interaction occurs, while ensuring accurate recognition in relevant segments.
Data Source
Figure 1
Figure 2
Figure 3(a)~3(c)
AI summary
In one example, a system for recognizing an object includes a processor to select from a plurality of image frames an image frame in which a view of the object is not blocked, and to estimate a location of the object in the selected image frame.