FAST AND ROBUST MODEL FOR MULTIPLE OBJECTS TRACKING USING KEY-FRAME DETECTION AND CO-TRAINED CLASSIFIER

Abstract

This paper proposes our new approach for multiple objects tracking for real-time video tracking applications. The new tracking method can improve tracking speed and reduce track fragmentation and confusion by using two convolutional neural networks  to detect and distinguish the targets. This mechanism ensures real-time capability when you do not have to perform deep learning detector continuously while still ensuring constant and accurate updating of the target's position. This is called a co-training mechanism. The keyframe detection model is a Single Shot Detector that also operates as a data generator; the second neural network is a classifier that will be trained from data collected from the main detector. The tracker is presented as a combination of techniques that we named DCT (Detector-Classifier Tracker). This article will fully explain the working mechanism of DCT and presents the test results for the combined image attachment method according to the frame processing experiments on data of long range thermal imaging cameras.