Object Tracker with Deep Reinforcement Learning with Application to Autonomous Driving

Abstract

Perception is an essential part of Self-driving Car or ADAS (Advanced Driver AssistanceSystem). Recently Object detection and Object tracking task have improved by CNN (convolutionalnerural network) and huge data set such as Imagenet and COCO. However, Performance of Objecttracking tasks has suffered from insufficient video sequence data set unlike a high resolution large dataset for object detection. In this work, we propose tracker model combined with deep reinforcementlearning for solving limitation of dataset quantity problem. Proposed model estimates next directionand position of bounding box using raw RGB image data. This model consider the learning process assemi-supervised learning rather than supervised learning. Using this approach, tracker can be learninsuffcient data efficiently. We also consider tracking problme as class-agnostic manner, it is moresuitable manner to avoid arbitrary dangerous obstacle in autonomous vehicle field. We proveavailability of proposed model using VOT2015, OTB-100 and our diving data.