Vision Based Object Counting Using Speeded Up Robust Features for Inventory Control

In this paper, an approach for industrial machine vision system is introduced for effective maintenance of inventory in order to minimize the production cost in supply chain network. The objective is to propose an efficient technique for object identification, localization, and report generation to monitor the inventory level in real time video stream based on the object appearance model. The appearance model is considered as visual signature by which individual object can be detected anywhere via camera feed. Herein, Speeded Up Robust Features (SURF) are used to identify the object. Firstly, SURF features are extracted from prototype image which refers to predefined template of individual objects, and then extracted from the camera feed of inventory i.e. scene image. Density based clustering on SURF points of prototype is done, followed by feature mapping of each cluster to SURF points in scene image. Homographic transforms are then used to obtain the location and mark the presence of objects in the scene image. Further, for better invariance to occlusion and faster computation, a novel method for tuning the hyper parameters of clustering is also proposed. The proposed methodology is found to be reliable and is able to give robust real time count of objects in inventory with invariance to scale, rotation and upto 70% of occlusion.