Research On Deep Reinforcement Learning Based Salient Object Detection And Tracking Method

Computer vision is one of the important research contents in the field of artificial intelligence.In recent years,visual methods have made good progress through their combination with in-depth intensive learning.Starting from the visual problems in complex scenes,this paper presents solutions for object detection,salient object detection and object tracking with deep reinforcement learning.The results show that this measure effectively improves the accuracy of detection and tracking.The main contributions of this paper are as follows:(1)To solve the computational cost problem caused by a large number of object proposals in the current object detection method,the object detection process is defined as a Markov decision-making process referring to the human vision mechanism of detecting a target in this paper.A detection agent with context comprehend and hierarchical actions are constructed.The number of candidate areas and the actions taken by the agent are very small with the top-down search strategy,which makes the computation dramatically reduced.In addition,fine actions further designed for the agent's search strategy make the location of an object more precise.Finally,the optimal strategy is obtained through trained the agent with reinforcement learning.(2)A two-stage salient object detection method is proposed to solve the problem of low detection accuracy of salience objects in complex scenes.The two phases contain the salient region localization and the salient object segmentation.Deep reinforcement learning is used to train the agent to locate the salient area step by step through executing a sequence of actions,and then the target is segmented finely.The feature extraction network is shared as the backbone network to simplify the model and reduce the number of parameters,and a divide-and-conquer training strategy is proposed for the two-stage algorithm.The experimental results show that the algorithm eliminates the interference information by salient region location,makes the segmentation results of salient objects more accurate,and has real-time detection performance.The results of pedestrian detection datasets show that has strong generalization ability on other practical application problems.(3)A stage-wise online tracker selecting method is proposed to solve the problem that a single tracking algorithm cannot adapt to the different challenges that occur in different time periods in complex scenes.Considering the incomplete perception of a frame of tracking scenes transformation,this method is defined as a partially observable Markov decision process.An agent is trained to select the optimal candidate tracker for a current frame with deep reinforcement learning.This work collects a large number of traffic videos and builds a dataset for training and testing.Experiments show that the method can improve the accuracy and robustness of the tracking algorithm through combine the advantages of candidate trackers.Tests on public datasets further demonstrate the potential and advantages of this approach.