Object Detection with Location Priors for Occluded Multi-View Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection systems suffer from missed detections when targets are heavily occluded, leading to increased false positives or higher costs for additional data collection and reduced model efficiency.
Innovation Solution
Utilize a location prior derived from a view with strong object detection to enhance detection in views with occluded objects by blending probability maps from multiple camera views or a joint tracker.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If object detection is performed using a single camera view, then the detection process is simple and fast, but missed detections increase when objects are heavily occluded
Solution Approach 1:
The patent combines probability maps from multiple camera views to create a comprehensive detection result. By merging the first probability map from the first camera view with the second probability map from the second camera view, the system achieves more reliable object detection while handling occlusions effectively.
Solution Approach 2:
The patent introduces a temporal dimension by capturing images at different time points and generating probability maps across multiple camera views. This multi-dimensional approach (spatial across cameras + temporal across frames) allows the system to track objects through occlusions and improve detection reliability.
2Reliability
If multiple camera views are used to improve detection accuracy, then occlusion problems are reduced, but the system complexity and processing cost increase
Solution Approach 1:
The system performs preliminary object detection and generates probability maps from multiple camera views before final object detection. By pre-processing images from both cameras and creating probability maps in advance, the system reduces the complexity of the final detection step while maintaining high accuracy.
Solution Approach 2:
The patent introduces probability maps as intermediary representations between raw images and final object detection results. These probability maps serve as mediators that consolidate information from multiple camera views, simplifying the integration process and reducing system complexity.
3Measurement precision
If traditional object detection methods are used, then the system is simple to implement, but false positives increase when targets are heavily occluded
Solution Approach 1:
The patent merges probability maps from multiple camera views to produce a blended probability map. This combination approach reduces false positives by cross-validating detections across different views, improving measurement precision while maintaining reasonable system complexity.
Data Source
AI summary
Techniques are described for object detection. For example, a computing device can: determine, based on a first image of a scene obtained from a camera with a first view of the scene, a first probability map including probabilities of object(s) being located at locations within the scene; determine a location of an item associated with each object in the first image; map the item from the first view to a second view to produce a prior probability map associated with the second view. The computing device can obtain, from the camera/another camera with a second view of the scene, a second image of the scene; determine, based on the second image, a second probability map including additional probabilities of the object(s) being located at the locations; blend the second probability map with the prior probability map; detect, based on the blended probability map, the object(s) of the scene.