Detection of Customer Interested Garments in Surveillance Video using Computer Vision

TL;DR

The paper addresses detecting garments a customer is interested in from surveillance footage in Indian retail settings. It implements a Mixture of Gaussians background subtraction pipeline with color-based segmentation and a YOLOv3 filter to isolate garment regions from customers and background. On a two-camera, 944×576 dataset, it reports an average IoU of $70.81\%$ and precision/recall of $0.82$ at a threshold $0.55$, demonstrating feasibility in noisy, low-resolution footage. This approach enables customer-interest analytics and potential sales optimization, with future work targeting crowd density effects and demand prediction.

Abstract

One of the basic requirements of humans is clothing and this approach aims to identify the garments selected by customer during shopping, from surveillance video. The existing approaches to detect garments were developed on western wear using datasets of western clothing. They do not address Indian garments due to the increased complexity. In this work, we propose a computer vision based framework to address this problem through video surveillance. The proposed framework uses the Mixture of Gaussians background subtraction algorithm to identify the foreground present in a video frame. The visual information present in this foreground is analysed using computer vision techniques such as image segmentation to detect the various garments, the customer is interested in. The framework was tested on a dataset, that comprises of CCTV videos from a garments store. When presented with raw surveillance footage, the proposed framework demonstrated its effectiveness in detecting the interest of customer in choosing their garments by achieving a high precision and recall.

Figures (5)

• Figure 1: Block diagram of the proposed garments detection approach
• Figure 2: Garments of interest as detected by the proposed approach. (Best viewed in colour)
• Figure 3: Some of the sales videos in the dataset.
• Figure 4: Variation of precision and recall with IoU threshold.
• Figure 5: The number of true positives, false positives and false negatives at IoU threshold of 0.55.