challenges in stock visibility, manual inventory counting, and inefficient resource utilization. The system

integrates artificial intelligence, computer vision, and image processing to enable real-time stock detection,

customer presence tracking, and parking lot traffic monitoring. Designed to enhance operational transparency

on manual supervision. The architecture comprises two core modules: an AI-based Video Recognition Module

related to hardware dependency and difficulty in detecting low-stock items under full occlusion. Future

enhancements will focus on cloud-based implementation, model optimization, and integration with point-of-sale

explainable, and scalable AI-driven framework that advances warehouse automation, improves resource

utilization, and strengthens real-time decision-making within retail environments.

Explainable AI, AI in Logistics.

Effective warehouse management is critical for modern retail operations, yet many warehouses continue to rely

delayed replenishment, and difficulty locating OOS areas. Additionally, monitoring in-store customer traffic and

parking availability is essential to optimize warehouse operations. Understanding traffic patterns allows

warehouse staff to allocate resources effectively, anticipate demand, and ensure that high-demand products are

replenished before shelves run empty. A lack of real-time traffic monitoring during peak periods or promotional

events may exacerbate temporary OOS conditions, affecting both sales and customer loyalty.

and image processing technologies, including YOLO (You Only Look Once) for object detection and SORT

(Simple, Online, and Realtime Tracking) for tracking and counting. The system provides a comprehensive,

automated solution for monitoring inventory, detecting OOS products, and analyzing in-store and parking traffic.

availability, and provision of actionable insights through an intuitive desktop interface. These capabilities allow

warehouse managers and staff to optimize resource allocation, expedite replenishment, and make informed

operational decisions. By enhancing inventory visibility, improving traffic monitoring, and automating repetitive

customer satisfaction. Its scalability and adaptability make it suitable for deployment across diverse retail and

warehouse environments, offering a sustainable and intelligent approach to modern warehouse management.

the challenges and digitization of inventory management, the application of computer vision in logistics, and the

evolution of real-time object detection models.

Studies have explored utilizing Radio-Frequency Identification (RFID) and sensor networks; however, these

technologies often require direct tagging of every item, presenting scalability and cost challenges for high-

volume, dynamic environments [2]. The limitations of conventional methods establish a clear requirement for

non-intrusive, vision-based alternatives.

Computer vision has emerged as a disruptive technology in supply chain optimization, offering solutions for

quality inspection, package sorting, and inventory assessment. Early implementations primarily used traditional

image processing techniques, but modern approaches leverage deep learning for superior performance under

complex conditions. Convolutional Neural Networks (CNNs) have facilitated significant breakthroughs,

classifying detection models into two main categories: two-stage detectors (e.g., R-CNN, Faster R-CNN) and

vastly improving speed while maintaining competitive accuracy.

TABLE I. YOLO Technology’s Evolution and Performance Analysis

YOLOv3 (2018)

Used Darknet-53 as the backbone and logistic classifiers for improved accuracy, achieving

an mAP of 28.2 at 22 milliseconds per image. This version continued to build on the speed

and accuracy improvements.

FPS on the COCO dataset. This version focused on enhancing the model's performance

without increasing computational cost.

YOLOv8 (2023)

Known for its high accuracy and speed, featuring a Python package and CLI-based

implementation for ease of use. YOLOv8 represents the latest advancements in the YOLO

family, integrating state-of-the-art techniques for superior performance.

Figure 1: YOLOv8 Performance Against Other Prominent Versions [7]

enhanced data augmentation, which were previously applied in logistics for monitoring assembly lines and

simple shelf checks [4]. However, these versions often required substantial computational resources, limiting

their deployment on edge devices or modest server infrastructure.

Advanced object counting and tracking techniques are critical for real-time applications, extending beyond

logistics into traffic management and public monitoring. Mandal and Adu-Gyamfi's study [8] provided a

comprehensive analysis of algorithms like YOLO, Faster R-CNN, SSD, SORT, and DeepSORT, concluding that

the combination of the YOLOv3 detector with the SORT tracking algorithm offers the optimal balance between

detection accuracy and real-time performance for vehicle counting applications. Building on tracking

performance, [9] proposed a novel scenario-based object counting method that introduces scene awareness by

integrating re-identification (Re-ID) with object detection (using CenterNet), specifically segmenting regions

and clustering targets to reduce ID swapping and improve accuracy in densely populated areas. Furthermore, the

problem of accurate counting in highly complex environments was addressed by [10], who introduced the

Locount task, which explicitly integrates object location and counting to robustly handle severe occlusion among

Effective real-time inventory and traffic management relies on a combination of robust image processing

techniques. This project's system integrates YOLOv8 for detection, SORT for tracking, and layer masking for

focused image processing. YOLOv8 is selected as the primary object detection algorithm due to its superior

speed and accuracy, evidenced by research showing a mAP50 score of 0.933 on retail products, ensuring fast

detection of items and traffic [12]. For sustained tracking across frames, the lightweight SORT algorithm is

employed; it uses the Kalman filter and the Hungarian algorithm to efficiently predict and associate object

motion, maintaining high accuracy even with fast movement and occlusion. Complementing these, layer

masking, supported by the work of [13], is crucial for enhancing robustness by selectively hiding or showing

parts of the frame to isolate noise and irrelevant details, e.g., masking external streets during car lot monitoring,

ensuring that detection and counting are highly precise and task-specific.

understanding, preparation, modeling, evaluation, and deployment. Objectives focus on automated front-face

product counting and OOS detection to enhance warehouse efficiency. Data collection included 935 images from

diverse shelf conditions, followed by augmentation (rotation, brightness, blur) and labeling via Roboflow, as

shown in Figure 2.

Launch stages incorporated real-world implementation and user training, allowing the system to adapt to

evolving requirements.

state of a linear dynamic system from a series of noisy measurements. It involves two main steps, i.e. the

prediction step to estimates the state at the next time step, and the update step to corrects the state estimate using

the measurement at the current time step. The state vector typically includes the position and velocity of the

object as in equation (2).

where

x, y represents the object's center coordinates

s is the scale area

r is the aspect ratio

𝑥󰇗, 𝑦󰇗, 𝑠󰇗 are the respective velocities

The Hungarian Algorithm is used for data association, solving the assignment problem to optimally match

predicted positions of tracked objects with new detections. This minimizes the total distance (measuring as cost)

between the predicted and the detected object positions.

In linear counting, defining a line and to check if the center point of the detected object crossing the particular

line is crucial to accurately counting each object only once as it passes a predefined virtual line. The key steps