Improvement And Implementation Of Self-Learning Visual Tracking Algorithm For Android

At present,the vision tracking algorithm is an important research direction of artificial intelligence,and a variety of visual tracking algorithms have been created.TLD vision tracking algorithm is one of them.Because of its combination of tracking and detection algorithm and online learning mechanism,the tracking effect of TLD algorithm is better than the mainstream vision tracking algorithm.Using the advantages of the portability of mobile devices,we can use TLD algorithm accurately track and analyze the targets anytime and anywhere.However,the mobile devices with ARM CPU(Central Processing Unit)is limited in processing capacity.Therefore,on the basis of TLD vision tracking algorithm,this paper proposes a Android based visual tracking algorithm.In this paper,the TLD algorithm proposed by Kalal Zdenek which includes tracking algorithm,detection algorithm and P-N learning mechanism are studied.To improve performance,the TLD vision tracking algorithm is analyzed.Because of the need to deal with too many scanning windows,the detection module takes the longest time.Therefore,three improvements are proposed to reduce the scanning window.In the implementation phase,considering of the efficiency of the system,the improved algorithm written with C++ is adapted to Android application,using the JNI(Java Native Interface)mechanism provided by Android.The improved algorithm is tested using video sequences in the Visual Tracker Benchmark with PC.Compared with the original TLD algorithm,the improved algorithm's processing time of each frame is improved by about 20%.The self-learning visual tracking system based on Android is tested using real object with smart phones.The system can track the object effectively,get the average frame rate of 12 fps(frame per second),and overcome the target deformation and occlusion the disappearance of the situation.The application meets the need for visual tracking in everyday life and meets design targets.