Active Object Detection Based on a Novel Deep Q-Learning Network and Long-Term Learning Strategy for the Service Robot

This article focuses on active object detection (AOD), one of the greatest challenges in the robotics field. A novel deep-Q-learning-network-based approach is proposed to utilize more useful status information for enhancing the training efficiency and testing accuracy of AOD by adding the cropped target object (TGOJ) from the current state as a new input. Different from the existing researches, a novel reward function, combing the area factor and distance factor of the bounding box, is designed to make the robot not only get closer to the TGOJ but also obtain a better observation viewpoint. Moreover, to overcome the differences between the training dataset and new environments as well as improving the adaptation of the AOD model, a reward-based long-term learning strategy including a novel training strategy is presented. The comparable experiments and the ablation study have been implemented in an AOD dataset, proving that our method owns better performance and efficiency than the comparable methods. Meanwhile, the experiments in the real-world scenario with a robot indicate the validity of the proposed method.