Implementation of Deep Learning-based Object Recognition and Tracking for Intelligent Video Surveillance

Abstract

As research on artificial intelligence technology is actively conducted in recent years, research on deep learning technology that recognizes and classifies images in real time on behalf of humans is required. Object recognition has difficulties in finding an object of interest from a video clip or image, and classifying several detected objects for each object. To solve this problem, research is needed to detect and track objects using a CNN-based deep learning method. Among the CNN-based multi objects detection techniques, the most well-known methods are R-CNN and Faster R-CNN. These methods are based on ROI-based detection techniques to perform verification work to reduce candidate groups through pre-work in the ROI of the object. However, since the process of classifying objects for each region of interest is performed, the detector speed decreases. Real-time processing is not possible due to this speed problem. In this paper, to overcome these issues, we have proposed multi object detection, classification, and tracking method using YOLO, a single step technique that performs a single CNN to determine the location and type of objects in an image. Experimental results depict that it can detect and classify objects robustly in various environments, and that real-time tracking is possible because the calculation speed is faster than the conventional method.