The retail and industrial landscapes are undergoing a digital transformation 2.0, where real-time analytics and advanced computer vision are driving significant advancements. Accurately counting unique individuals or distinct objects in dynamic environments such as shoppers walking through a store, browsing products on shelves, or workers in a factory can improve operational efficiency and customer experience. This is powered by AI-driven solutions that leverage deep learning models to analyze video streams in real-time, ensuring precise object detection, tracking, and identification even in crowded or fast-moving scenarios.

 

By running advanced computer vision models for person or object detection and individual tracking on SiMa.ai Machine Learning System-on-Chip (MLSoC), companies can deploy real-time counting systems. These systems leverage SiMa.ai edge computing to accelerate low-latency inference with cloud-based post-processing for advanced analytics of an end-to-end application.

 

Real-world Impact: Enhancing Operations and Customer Experience

 

In fast-paced retail and industrial environments, real-time detection and tracking are crucial. SiMa.ai MLSoC edge compute devices power AI models that can process video streams locally, reducing latency and ensuring immediate insights.

 

Retail Footfall Analytics: Accurate object counting helps retailers optimize store layouts, improve staffing decisions, and measure marketing effectiveness. By understanding customer flow patterns in their stores, retailers can create more engaging and efficient shopping experiences.

 

Industrial Worker Safety: In manufacturing environments, real-time object counting ensures compliance with safety regulations and monitors worker movement in hazardous zones.

 

Driving Real-time Insights with SiMa Edge AI

 

Key components required for Real-Time Retail Zone analytics are:

• Visitor Detection: Detects unique visits in video frames with high accuracy using specially trained models such as CenterNet, Yolo class models, or Vision Transformers 
• Visitor Tracking: Associate objects across different frames or camera views by extracting feature embeddings using ReID models and enable multiple object tracking by using the “tracking by detection” approach which combines motion prediction and appearance matching.

 

System Workflow includes the following steps:

       1. Video Capture: GigE cameras RTSP stream live video feeds to MLSoC edge devices.

       2. MLSoC Edge Inference: CenterNet detects people, and the ReID model generates feature embeddings. 

       3. MLSoC Edge Tracker: The tracking algorithm combines Linear Kalman Filter (LKF)  for motion prediction and feature vector comparison for appearance matching, associating current detections with existing tracks, while managing track creation, updates, and deletions to maintain consistent object identities across video frames.

      4. Cloud Processing: Embedding metadata is sent to the cloud for aggregation, deduplication, and advanced analytics.

 

The end-to-end pipeline uses CenterNet to detect people, and ReID to track people, followed by a tracker module that provides the unique ID based on the foot traffic within a user-configured ROI (Region Of Interest) zone. Further analytics and additional business logic done on the cloud with the real-time edge inference plus metadata which includes unique ID, timestamp etc. generated from the pipeline running on the MLSoC. See Fig.1 which highlights all the system modules and workflow configurations for the GStreamer based end-to-end pipeline implementation. The ingest begins with the RTSP camera source, followed by frame decoding, visitor detection using CenterNet, re-identification using ReID, tracking, result overlay, frame encoding, and finally streaming to a UDP host sink for efficient deployment.

 

Fig. 1 Real-Time Retail Zone Analytics Pipeline

 

Technical Details

 

For access to the complete source code and GStreamer pipeline implementation details, visit SiMa.ai developer guide. Table 1. highlights parameters defined in tuning_config.json file to help with fine-tuning to customer requirements. 

 

Table 1. Tuning Parameters

 

 

Key Performance Indicators

 

Table 2. Highlights a few metrics associated with the models and pipeline used for the real-time retail zone analytics application.

 

Table 2. KPI Metrics

 

Advanced Cloud Analytics for Enhanced Insights

 

Once data is processed on the edge, metadata such as bounding boxes, embeddings are sent to the cloud for advanced deduplication and aggregation.

 

Capabilities:

• Heatmaps: Identify high-traffic areas in retail stores or monitor crowded zones in factories.
• Anomaly Detection: Detects unusual patterns, such as unauthorized access to restricted zones.
• Predictive Analytics: Forecast footfall trends to optimize staffing and resource allocation.

 

Conclusion: Transforming Industries with SiMa ONE Platform Edge AI

 

The combination of real-time edge detection using CenterNet, ReID and sophisticated cloud analytics transforms how retail and industrial sectors count and track individuals. These AI solutions improve operational efficiency, enhance safety, and create more personalized customer experiences based on the in-store pilot implementation analysis.

 

Retail Chain Optimization:
By integrating CenterNet and ReID for unique people counting, we expect a ~20% improvement in customer flow efficiency and a ~15% increase in sales through better resource allocation.

 

Industrial Safety Compliance:
A manufacturing plant with AI-powered tracking to monitor worker safety in hazardous areas, we expect to reduce safety incidents by ~30% and improve compliance with regulatory standards.

 

Discover how SiMa.ai can tailor an AI solution to meet your unique needs and drive transformative outcomes.

Join our Edgematic Early Access Program to experience our No Code/Low Code, Drag & Drop Programming Environment, empowering you to create, build, and deploy edge AI solutions seamlessly on SiMa.ai MLSoC silicon. Contact our solutions experts today.