Partial Inference Path Neural Network for Video Object Tracking

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current machine learning technologies, such as deep neural networks, face high computation costs and latency issues in object tracking, especially when dealing with multiple objects, and often suffer from quality problems like drifting and ghosting.

Innovation Solution

Implementing a partial inference path in neural networks for video processing, which uses only early feature layers and ROI pooling, allowing for efficient object tracking by leveraging temporal and spatial coherence between frames, reducing computational overhead and maintaining detection accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a full DNN inference path is used for object detection and tracking in video frames, then detection accuracy is maintained, but computation cost and latency increase significantly

Engineering Contradiction:

Improvedetection accuracyVSAvoidcomputation cost

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent segments the DNN inference process into two distinct paths: a full inference path for initial object detection and a partial inference path for subsequent object tracking. The partial path reuses features from earlier layers and skips later processing stages, thereby reducing computation cost while maintaining tracking accuracy through temporal coherence exploitation.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent implements a partial inference path that performs only a subset of the full DNN inference operations. Specifically, it reuses features from early layers and performs only necessary processing for tracking, avoiding redundant computations while maintaining sufficient accuracy for tracking scenarios where objects are already detected.

Inventive Principle:

Principle #16Partial or excessive action

2Adaptability or versatility

If object-agnostic trackers are used to track multiple objects, then tracking coverage is improved, but tracking cost exceeds detection cost and quality issues arise

Engineering Contradiction:

Improvetracking coverageVSAvoidtracking cost

Core Design Contradiction:

Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent performs preliminary object detection using the full inference path to obtain initial object locations and features. Subsequent tracking operations then leverage these pre-detected objects and their associated features, avoiding the need for expensive re-detection while maintaining accurate tracking of multiple objects throughout the video sequence.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS11699063B2Partial inference path technology in general object detection networks for efficient video processing

Publication Date: 2023.07.11 INTEL CORP

AI summary

Systems, apparatuses and methods may provide for technology that generates, by a full inference path of a neural network, a first detection result associated with one or more objects in a first video frame. The technology may also generate, by a partial inference path of the neural network, a second detection result based on the first detection result, wherein the second detection result corresponds to a second video frame that is subsequent to the first video frame.