The client operates a network of over 100 supermarkets, providing a wide range of consumer goods. Their offerings include groceries, household items, and personal care products, catering to the daily needs of the average consumer.
In the retail sector, supermarkets are focusing on digitalization to meet evolving customer expectations and maintain competitiveness. They are challenged with improving customer experience by precisely understanding consumer behavior, enhancing operational efficiency, and employing data analytics for strategic decision-making and planning.
The client's adoption of SciForce’s EyeAI 👁️ is a strategic move in their digital transformation. It serves to enhance customer insights through advanced video analytics and improve store management by optimizing customer flow and queue dynamics.
EyeAI addresses the challenge of efficiently optimizing space and managing queues in sectors like HoReCa, retail, and healthcare. It provides real-time insights into visitor behavior using existing camera systems, eliminating the need for new hardware. This makes it a cost-effective solution for businesses seeking to improve operational efficiency and customer service.
1. Inefficient Space Utilization
Businesses face challenges in optimizing their space, affecting customer experience and sales. EyeAI provides crucial insights for efficient layout planning, essential in sectors like retail and hospitality where product or service placement greatly impacts customer engagement and revenue.
2. Ineffective Queue Management
Businesses often struggle with long customer queues, leading to dissatisfaction and lost sales. EyeAI improves queue management by providing real-time analytics on customer flow, enabling proactive adjustments in staffing and service, reducing wait times and increasing efficiency.
3. Lack of Real-time Visitor Data
Businesses in customer-focused sectors often lack immediate data on visitor behavior, missing chances for service personalization and customer responsiveness. EyeAI addresses this by providing real-time behavioral analytics, enabling businesses to quickly adapt to customer preferences and enhance relationships and loyalty.
EyeAI leverages AI to analyze video feeds from existing cameras, specifically targeting space optimization and queue efficiency in customer-focused environments. This system identifies patterns in visitor behavior and crowd dynamics, offering real-time insights for businesses to enhance layout planning, manage queues effectively, and personalize customer interactions.
Key Features
1. Identify & Analyze Visitors
Detects visitor behavior patterns, such as shopping habits and time spent in different areas. Allows businesses to offer personalized offers and recommendations, based on specific customer interactions.
2. Space Utilization Efficiency Analysis
Evaluates occupancy and facility usage data to optimize space efficiency and flow. It helps to improve overall space usage efficiency, improve customer experience, and increase sales.
3. Visitor's Trajectories Sampling
Tracks real-time visitor routes based on camera data. Provides insights on product placement optimization based on customer behavior, increasing the efficiency of merchandising strategies.
4. Entry & Exit Time Monitoring
Monitors the time customers enter and leave, providing a detailed analysis of peak traffic periods, average shopping duration, and patterns in customer flow. This data helps in staff scheduling, store operation planning, and creating personalized offers.
5. Queue Progress Management
Analyzes queue length and movement speed, along with crowd size in waiting areas. It also evaluates transaction speed and customer interaction in checkout areas, providing insights to improve service quality at checkout points.
6. Popular & Ignored Zone Detection
Identifies high-traffic and ignored zones, assisting in space rearrangement, product placement, and promotional displays.
These features collect all the information from existing cameras, giving a deep understanding of customer behaviors and space usage. Businesses get data-driven insights, allowing precisely adjusted service delivery, and marketing strategies, and use the space effectively, improving overall efficiency and customer service.
EyeAI operates only with existing camera data, without the need to purchase extra costly hardware.
Conceptualization
1. Challenge 1: Two-Step Detection Model with Verification
End-to-end person detection models work well for image analysis but have weaknesses when it comes to video processing. They often come up with false positives and imprecise bounding boxes. To address these issues, we introduced an advanced two-step person detection and verification model.
Detection phase The person detection phase implements the end-to-end YOLOv3 model, generating precise rectangles around detected individuals. We chose YOLOv3 for its optimal balance of accuracy and processing speed.
Verification phase The verification phase is key to improving detection accuracy. Using rectangles extracted from the detection module, the system is capable of filtering out false-positive samples. This step is vital for clean data collection, minimizing errors in ongoing analysis.
Model Development and Adaptation
2. Challenge 2 Enhancing Detection in Low-Resolution Environments
The low-resolution camera challenge involves accurately detecting individuals with cameras that produce images lacking in clarity and detail, affecting analytics accuracy. To tackle low-resolution camera challenges, EyeAI introduced an innovative method by creating "skeletal" representations from initial detections, using the advanced PoseNet and InceptionV3 neural networks.
It was noted that background noise levels are linked to the detection module's prediction confidence, as detected outlines may include irrelevant pixels. To address this, we applied a verification method to clean the person’s representation on a frame from the redundant background pixels. This enhances the accuracy of determining the customer's position in queue zones in service areas.
The processed frame, through the described pipeline, extracts "skeletal" key points of detected people, reducing background noise and false positives. These values are then inputted into the high-level model, which predicts if it is a queue or non-queue situation.
3. Challenge 3: Improving Accuracy with Verification and Skeletal Representation
Given the prevalence of low-resolution cameras, the detection pipeline can vary. We use classical computer vision techniques for detecting humans and other moving objects effectively.
Image processing The image is processed through morphological operations such as dilation and erosion. As a result, you have a grayscale-channeled mask, with the removed noise and restored objects of interest.
Object tracking with GSRT tracker We get the data for moving object tracking processing of the video frames with the CSRT tracker. We chose it for its high accuracy and reasonable computational demands, allowing for efficient tracking of moving objects.
After getting the list of the moving objects, the custom-written post-processing filters out the redundant objects based on their size. The model evaluates the average size of the box to capture the most frequently moving type of object (humans in our case). We also need to take into account camera positions and increasing distance between the camera and the object
Verification and Enhancement For the verification step, PostNet can be used, noting that images might not always capture detailed skeletons. To address this, super-resolution models are recommended to enhance image quality and detail.
Super-resolution models, especially those based on GANs, are preferred for enhancing barely visible silhouettes by adding necessary context. Traditional interpolation methods, such as bicubic or area, are less effective and not recommended for use in high-reliability systems.
Model Enhancement & Implementation
High-level model
The high-level model in our system predicts queue presence using data from the low-level models we described above. Research shows that a linear machine learning model is ideal for this task because it's both complex enough to handle the task and fast in execution. This speed allows us to create customized queue predictions for each service shop. To further tailor these predictions, we introduce an online learning feature. This feature enables users to specify which areas should be considered as queues. The system then adjusts and improves its predictions based on these inputs, ensuring the model is finely tuned to meet the specific needs of each location.
Camera processing algorithm
The camera processing algorithm uses a scheduler to manage how data moves from cameras to both low and high-level models efficiently. This scheduler consistently checks and updates camera settings, like the web address, area coverage, and how fast it processes images. Since we use multiple cameras that work at different speeds, this system is designed to handle data asynchronously, ensuring smooth operation across all camera feeds.
Final Testing, Deployment, and Training
To guarantee EyeAI's accuracy and performance, we undertook comprehensive testing and adjustments. Following its successful validation, we tailored EyeAI’s deployment to meet the diverse requirements of each client environment.
We provided detailed training on EyeAI's features, usage, and best practices. A support framework was also set up to quickly address any operational questions or issues.
By launching EyeAI, we propose a new way to make the most of the camera systems already used in retail, healthcare, hospitality, and public safety sectors:
Impact on industry
EyeAI is an easy-to-use and cost-effective solution for improving commercial space management and queue optimization. It offers targeted insights for optimizing store layouts and minimizing customer wait times. The product is cost-effective and gives immediate results since it doesn’t require additional hardware installation and works with real-time data. EyeAI's technology has raised operational efficiency and customer service standards, steering the industry towards adaptive, data-driven strategies.
Impact on client \ end-user
After implementing EyeAI, the supermarket chain experienced optimized store layouts and effectively managed queues, leading to a faster, more pleasant shopping environment for customers. Shoppers benefited from reduced wait times and a more enjoyable visit. Overall result – efficient space usage, streamlined queue processing, and increased customer satisfaction and loyalty.
We are proud to solve real-world challenges in retail – crowded spaces and long queues are no problem anymore. EyeAI 👁️ has already proved its effectiveness.