During this time of rapid change in retail’s structure and consumers desire to upgrade, a lot of China’s retail landscape is changing from being focused solely on products to now include also a focus on experiences. However, retail managers in the past have primarily used manual means (like purely observing how well things are selling) and after-the-fact surveys to collect data about how their stores and/or products are performing. These methods suffer from many weaknesses, including an element of "disconnect" due to having multiple different ways that retail managers collect data on how their stores are doing across different stores, time periods, too much being required to process and interpret raw data, and not being able to identify potential areas for improvement (so called "churn points") within their retail environments. As a result of these issues, there is an immediate need to implement a comprehensive intelligent system that will provide data, analysis, and decision support to enable retail managers to improve their stores’ effectiveness. This paper presents a computer vision based customer flow analysis and service optimization model that is applicable to offline retail settings. The primary goals of the model are to increase customer satisfaction, improve the layout of products and increase the efficiency of service. For trajectory acquisition, the YOLOv8n object detection technology is combined with ByteTrack multi-object tracking technology. Identity association across cameras was accomplished through the OSNet re-identification network and each customer’s trajectory sequence is standardized (has a common format) and includes the coordinates of the customer’s location, the time of the customer’s position, the speed of the customer’s travel and the direction of the customer’s travel. In addition, complex customer behaviors such as pausing, wandering and hesitating will be analyzed using the SlowFast dual-path network and a customer-product gaze correlation matrix will be created to quantify the ability of a customer to fix their gaze on an SKU. Kernel density estimation will be used to create heat maps and divide the heat maps into hot, warm and cold zones. A Path Complexity Index (PCI) will be used to evaluate the efficiency of customer movement, and a four quadrant model of sales per square meter and attention level will be created in order to evaluate the performance of shelves. Finally, a graded intervention mechanism for hesitation behavior and a dynamic service resource scheduling strategy will be implemented to create a management closed loop of ‘perception-analysis-decision-optimization’. The model was assessed in an empirical study in a chain convenience store with 2,847 valid trajectory data points collected over a 14-day period. The results showed that the average customer PCI was 1.82; the seven "hot zones" (stores occupying only 18.5% of the area) handled 47.3% of customer traffic, achieving a sales per square meter (SPM) of 2.56; while the "cold zones" (stores occupying 46.3% of the area) had an SPM of only 0.30. The four-quadrant model identified 35.4% of problematic products and 18.7% of potential products. The conversion rate of wandering customers increased by 39.3%, the average waiting time decreased from 3.2 minutes to 2.1 minutes (-34.4%), and customer satisfaction increased from 3.6 to 4.2 (+16.7%). This model effectively breaks down the "data black box" of offline retail customer behavior, transforming computer vision technology into a practical management decision-making tool, providing theoretical support and a practical paradigm for retail enterprises to achieve low-cost, high-efficiency refined operations and experience upgrades.
Haoran Chen (Thu,) studied this question.