Autonomous Object Detection And Tracking Algorithm

The autonomous object detection and tracking technology is one of the core technologies of the aircraft to achieve object detection and search.It can lock and track the object quickly during cruise or reconnaissance missions and play an important role in analyzing the battlefield situation and object state in real-time.In this thesis,an autonomous object detection and tracking scheme is proposed to solve the problem of inaccurate object initial position during the human jitter or rapid object movement in aircraft cruise and reconnaissance missions.The object detection module automatically selects the object and provides the object initial position.The object tracking module performs object tracking according to the information from detection module.In the object detection module,the thesis discussess convolution neural networks such as R-CNN,YOLO and SSD.Considering that the traditional method is computationally intensive and the features of manual design are not portability,the thesis introduces the SSD network into the object detection module and fine tunes the SSD network.At the same time,since that the output of object detection in real scene is multi-class,the priority ranking strategy is proposed and designed.The SSD-based object detection algorithm is performed to provide a single reliable location for the object tracking module.In the object tracking module,this dissertation proposes a novel algorithm based on fusion feature and combines the correlation filter with the color probabilistic feature to improve the performace for the object deformation and fast motion.For the multi-scale change of the object,the thesis puts forward to merge the position and scale into one template and designs seven different scales for global optimal search.This thesis designs a template updating strategy and modifies the learning rate according to the different confidence(peak energy)of the response map.Meanwhile,put forward the idea of cycle updating to reduce the calculation complexity.This dissertation analyzes the various object features and introduces the convolutional feature of the object detection module into the tracking module.On the basis of confidence value(peak energy)of response map,this thesis also designs a re-detection scheme for tracking drift to catch the goal of autonomous object detection and tracking.