Modelled Neural Networks for Multiple Object Tracking

Recent development in artificial intelligence brought deep learning and neural networks that are applied in various areas, e.g. robotics, surveillance, autonomous driving, automation and medicine. Multiple Object Tracking very commonly utilises those architectures and there are many different approaches for this task. Those solutions are based on different kinds of neural network structures and this paper provides comparison of the corresponding algorithms that could improve further research. The paper investigates the performance of the Faster R-CNN, the VITAL and the RetinaNet methods with practical results and examines their different architectures used for object detection. The requirements for models are the detection of objects’ position and their classification. For tracking the instances, we use algorithms that are based on object detection systems. For registering the location of items Neural Networks use the IOU (Intersection of Union) in order to determine which bounding boxes should be examined and according to the IOU we distinguish positive and negative proposed bounding boxes. The negative predictions impact the performance and negatively contribute to the wanted signal. The results of the Faster R-CNN method present those challenges. The object classification could become difficult in the event of occlusion. The RetinaNet method provides distinguished detection and classification results that could be applied for the Faster R-CNN and the VITAL computations. There are many evolving implementations for object tracking. The VITAL detector that uses the GAN for the motion prediction was evaluated on the custom set of image sequences, that are used for deep neural network adaptive parameter regulation..