Pedestrian Video Tracking for Accurate Shelf Attention Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining the attention degree paid to store shelves rely on manual observation or inferred sales data, which are resource-intensive, prone to errors, and incapable of providing accurate, scalable data for optimizing shelf layouts and product selection.
Innovation Solution
Utilizing deep-learning models to analyze video frames from a video capturing equipment, detect pedestrians, generate bounding boxes, identify feature information, and record pedestrian historical data in a database, with deduplication processes to reduce redundancy and enhance data efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual observation methods are used to determine shelf attention degree, then implementation simplicity is maintained, but measurement precision and data reliability are insufficient
Solution Approach 1:
The patent replaces manual observation (mechanical/human system) with automated video capture and deep learning analysis (optical/computational system). Video capturing equipment automatically records customer behavior, and deep learning models process the visual data to determine attention degree, eliminating manual intervention while significantly improving measurement precision.
Solution Approach 2:
The patent creates a digital copy of customer behavior through video frames and processed images. Instead of directly observing customers, the system captures visual information, converts it into digital representations (bounding boxes, feature extractions), and analyzes these copies to determine attention degree, enabling precise automated measurement.
2Measurement precision
If sales data is used to infer shelf attention degree, then data collection is simplified, but measurement precision and reliability deteriorate due to indirect inference
Solution Approach 1:
The patent replaces indirect inference from sales data with direct observation through video analysis. The deep learning models directly analyze customer behavior visual evidence (gazing, approaching, interacting with products) to determine attention degree, providing direct measurement rather than indirect inference, thus improving precision while maintaining efficient processing through automated algorithms.
3Measurement precision
If comprehensive pedestrian data is collected for accurate attention analysis, then measurement precision improves, but data redundancy and processing complexity increase
Solution Approach 1:
The patent extracts only the necessary feature information from video frames using deep learning models. Instead of storing and processing all raw video data, the system extracts key features (pedestrian attributes, behavior patterns, bounding box coordinates, attention metrics) and stores only these extracted features, eliminating data redundancy while maintaining analysis precision.
Solution Approach 2:
The patent creates simplified digital representations (copies) of pedestrian data. Instead of storing original video frames, the system generates compressed feature vectors, bounding box annotations, and behavior classification labels that capture essential information in a compact format, reducing storage requirements while preserving analytical value.
Data Source
AI summary
The present disclosure provides an information acquiring method and apparatus, and an attention degree detecting system. The information acquiring method includes: acquiring a video frame captured by a video capturing equipment detecting whether a pedestrian exists in the video frame by using a first deep-learning model; processing the video frame by using a second deep-learning model to generate a pedestrian bounding box of the pedestrian under a condition that the pedestrian exists in the video frame, and assigning a corresponding pedestrian identifier to the pedestrian bounding box; identifying an image in the pedestrian bounding box by using a third deep-learning model to identify feature information of the pedestrian; and writing a pedestrian historical data into a database, wherein the pedestrian historical data comprises a timestamp and an identifier of the video frame, coordinate information of the pedestrian bounding box, the pedestrian identifier and the feature information of the pedestrian.