Retail analytics is needed for effective store management. Big stores have the requirement to track the customer behaviour in order to provide best service to them and also to improve their products so that they remain the first choice for the consumers. The client was a grocery chain that wanted to understand consumer behaviour from their CCTV video recordings. They wanted us to build a system that can analyze the store aisles recordings to assess the consumer behaviour based on some metrics. For example, the customer footfall metric helps them to equip their store with an appropriate number of support staff to service their customers. In order to build the system, we were required to study shoppers with metrics namely, gender, age, products of interest according to their age to build a profile for the shoppers frequenting the store. These metrics help estimate the customer footfall, consumer type estimation, dwell time etc.
This project involves integration of various Computer Vision algorithms namely person detection, face emotion recognition, age - gender detection and custom object detection. Of all the above said algorithms, only the custom object detection of retail stack of consumables required custom training. For the rest of the algorithms we integrated pre-built models from OpenVINO toolkit from Intel. The data acquisition phase of the project involved collecting images pertaining to the retail stack of consumables under varying lighting conditions and camera noise conditions. This was then followed by an object annotation phase with the CVAT tool to generate a segmentation dataset of 2000 images. We implemented a neural network training process derived from the CNN architectures including the Yolo, Xception and Inception. From our RoC studies, it was observed that the Yolo V2 Lite model was performing adequately in terms of both accuracy and execution time. We then developed a CCTV video stream processing application that integrates the output from the various object detection algorithms and implemented a robust reporting visualization program. The video stream processing server was then deployed as an on-premise installation that integrated with their existing CCTV NVR system.
We are able to introduce a novel retail analytics process for their grocery store operation which gave unique insights about consumer interests and buying patterns. This project was developed in a time frame of 30 weeks. The CCTV video analytics server was able give insights to our clients about ideal grocery store layout and manpower scheduling. They were able to attract 15% more shoppers by making layout and manpower changes to their grocery chain.