Deep Learning Retail Tracking Without Kalman Filters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for tracking movement within large confined areas, such as stores, are ineffective due to reliance on algorithms like Kalman filters and SORT techniques, which are complex and unsuitable for real-time employee and shopper tracking, especially in environments with multiple cameras and varying entry/exit points.
Innovation Solution
A modified deep SORT technique that removes the Kalman filter and uses convolutional neural networks to identify individuals and track their movement, creating and deleting person identifications only at specific entry and exit points, allowing for accurate tracking without spatial camera coordinates and enabling multiple camera use.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional algorithms like Kalman filters and SORT techniques are used for tracking, then tracking can be performed, but the system becomes complex and ineffective for large-scale confined areas
Solution Approach 1:
The patent extracts and removes the Kalman filter component from the traditional SORT algorithm, creating a simplified tracking system that relies on detection results without complex prediction mechanisms. This extraction resolves the contradiction by eliminating the source of complexity while maintaining tracking functionality through direct association of detection results across frames.
Solution Approach 2:
The patent segments the tracking process into distinct components: detection, identification, and association. By separating these functions and using simple distance-based association instead of complex Kalman filter predictions, the system achieves effective tracking in large-scale areas without the computational burden of traditional algorithms.
2Measurement precision
If velocity-based prediction algorithms are used, then movement prediction can be achieved, but they become ineffective for large-scale confined areas
Solution Approach 1:
The patent changes the fundamental parameters used for tracking from velocity-based predictions to distance-based associations. By measuring and comparing positions directly across frames using simple Euclidean distance metrics, the system achieves accurate tracking in large-scale confined areas without relying on velocity predictions that fail at scale.
3Area of stationary object
If multiple cameras are used for tracking, then coverage of large areas is improved, but spatial coordinates and camera alignment become complex
Solution Approach 1:
The patent creates a universal tracking framework that works across multiple cameras without requiring complex spatial coordination. The system uses a unified distance-based association method that can handle detections from any camera independently, allowing multiple cameras to contribute to tracking without needing precise spatial relationships or alignment between them.
4Productivity
If real-time tracking is implemented, then operational efficiency is improved, but computational complexity increases
Solution Approach 1:
The patent employs simple, computationally inexpensive tracking objects that can be quickly created and discarded each frame. By using basic distance calculations and simple association rules instead of complex Kalman filters, the system achieves real-time processing capability with minimal computational overhead, making tracking feasible at practical frame rates.
Data Source
AI summary
The system and method discussed herein can capture images from one or more video streams of a store area, can use deep learning to identify people in the images as being a store employee or a shopper, and can use the deep learning to track movement of the people within the store. The tracked movement can provide information that is useful to operators of the store, such as where store employees are, how long they have been in certain areas of the store, which areas of the store need more employees, where most shoppers are concentrated within the store, which areas of the store are popular, and so forth. The system and method can provide instructions to employees on mobile devices or kiosks, in response to the employee locations and activity in the store area. The system and method can also log the movement information, for downstream use.