Multi-object Tracking via Generic Proposals and Particle Filters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Multi-Object Tracking (MOT) algorithms face challenges in efficiently detecting and tracking specific objects in video streams due to the computational cost of exhaustive sliding window searches and the reliance on general object detection methods, which are not adapted for tracking.
Innovation Solution
The method employs generic object proposals to predict candidate object locations, extracts multi-dimensional feature representations, and uses a particle filter for tracking, reducing the number of hypotheses and leveraging more complex models for improved accuracy and scalability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If exhaustive sliding window searches are used to detect objects at multiple scales and viewpoints, then detection accuracy is improved, but computational cost increases significantly
Solution Approach 1:
The patent segments the exhaustive search space into two parts: (1) a limited set of candidate windows generated by generic object proposal methods that cover most objects in the image, and (2) the remaining search space that is pruned or processed differently. This segmentation allows the system to focus computational resources on the most promising candidate regions while avoiding unnecessary computations in less relevant areas, thereby maintaining detection accuracy while reducing overall computational cost.
Solution Approach 2:
The patent applies preliminary action by generating generic object proposals before performing category-specific detection. These proposals pre-identify candidate regions that are likely to contain objects based on general object properties such as contours and shapes. By performing this preliminary filtering step, the system reduces the number of candidate windows that require expensive category-specific classification, thus reducing computational cost while preserving detection accuracy for actual objects.
2Productivity
If a limited set of generic object location proposals is used to reduce computational cost, then processing speed is improved, but tracking adaptation is worsened
Solution Approach 1:
The patent makes generic object proposals multi-functional by using them for both detection and tracking purposes. The same set of candidate windows generated by generic object proposal methods is reused across multiple frames and for multiple object categories, rather than generating separate candidate sets for each. This universal approach maintains processing speed while enabling tracking adaptation through temporal consistency and appearance model updates that work across different categories and frames.
Solution Approach 2:
The patent incorporates feedback mechanisms where detection results and appearance models from previous frames inform the tracking process. The system uses feedback from detected objects to update appearance models and adjust candidate window selection in subsequent frames, allowing the limited set of generic proposals to adapt to specific tracking scenarios over time while maintaining high processing speed.
Data Source
AI summary
A tracking system and method are suited to tracking multiple of objects of different categories in a video sequence. A sequence of video frames is received and a set of windows is extracted from each frame in turn, based on a computed probability that the respective window contains an object, without reference to any specific category of object. For each of these windows, a feature representation is extracted. A trained detector for a selected category detects windows that constitute targets in that category, based on the respective feature representations. More than one detector can be used when there is more than one category of objects to be tracked. A target-specific appearance model is generated for each of the targets (e.g., learned or updated, if the target is present in a prior frame). The detected targets are tracked over one or more subsequent frames based on the target-specific appearance models of the targets.