Action Detection Using Spatiotemporal Saliency Maps

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current action detection technologies, such as bag-of-features action detectors, are not reliable for detecting complex actions like exchanging items, burying, or hauling due to short duration and interference from irrelevant actions or background clutter, leading to low detection accuracy.

Innovation Solution

The introduction of a spatiotemporal saliency map that weights features by their relevance for the specific action, improving the selectivity of the feature representation and enhancing the detection accuracy by emphasizing relevant features while reducing the impact of irrelevant ones.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If bag-of-features action detectors are used for detecting complex actions, then the detection process is simple and computationally efficient, but the detection accuracy is low due to interference from irrelevant actions and background clutter

Engineering Contradiction:
Improvedetection accuracyVSAvoiddetector complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the feature histogram into multiple bins, each representing different spatiotemporal characteristics of actions. By dividing the feature space into discrete bins and applying specific weighting to each bin, the detector can selectively emphasize features relevant to complex actions while suppressing irrelevant background features, thereby improving detection accuracy without requiring a completely new detector architecture

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by assigning different weights to different bins in the feature histogram based on their relevance to the target action. This allows the detector to treat different feature regions differently - enhancing important local features while suppressing less important ones, which improves detection reliability for complex actions without increasing overall system complexity

Inventive Principle:
Principle #3Local quality

2Reliability

If standard bag-of-features detection is used, then the implementation is straightforward, but the detection reliability is low for actions with short duration and subtle motion

Engineering Contradiction:
Improvedetection reliabilityVSAvoidease of implementation
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The patent changes parameters by introducing bin-specific weighting factors that modify the contribution of each feature bin to the overall detection score. This parameter adjustment allows the detector to become more sensitive to subtle motion patterns and short-duration actions by emphasizing bins that capture these characteristics, while maintaining the same underlying bag-of-features detection framework

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent performs preliminary action by pre-computing and storing optimal bin weights based on training data before actual detection. This preliminary weighting preparation allows the detector to automatically emphasize relevant features during detection without requiring complex real-time computations, thus improving reliability for subtle actions while keeping implementation straightforward

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP2915102B1Detection of human actions from video data
Publication Date: 2019.03.13 NEDERLANDSE ORG VOOR TOEGEPAST NATUURWETENSCHAPPELIJK ONDERZOEK TNO
  • EP2915102B1 patent drawingFigure 1
  • EP2915102B1 patent drawingFigure 2
  • EP2915102B1 patent drawingFigure 3

AI summary

A video action detection system uses feature a data extractor to extract feature data from video data at detected spatiotemporal interest points. A feature data quantizer assigns the extracted feature values to bins of a feature vector. Bin values are computed from a sum of contributions of spatiotemporal points of interest that have been assigned to the bin, with a bin dependent adjustment of a size of the sum and/or the contributions. The video action detection system computes a sum of match scores between the feature vector and reference vectors for the predetermined type of action. The bin dependent adjustment of the size of the sum and/or the contributions is adapted in a training step, based on partial match scores for individual bins obtained using an initial action detector.