In the context of air-ground robot collaboration, the apperance of the moving ground object will change greatly from the perspective of the drone and traditional object tracking algorithms can hardly accomplish target tracking in such scenarios. In order to solve this problem, based on the Parallel Tracking And Detection (PTAD) framework and deep learning, an object detecting and tracking algorithm was proposed. Firstly, the Single Shot MultiBox detector (SSD) object detection algorithm based on Convolutional Neural Network (CNN) was used as the detector in the PTAD framework to process the keyframe to obtain the object information and provide it to the tracker. Secondly, the detector and tracker processed image frames in parallel and calculated the overlap between the detection and tracking results and the confidence level of the tracking results. Finally, the proposed algorithm determined whether the tracker or detector need to be updated according to the tracking or detection status, and realized real-time tracking of the object in image frames. Based on the comparison with the original algorithm of the PTAD on video sequences captured from the perspective of the drone, the experimental results show that the performance of the proposed algorithm is better than that of the best algorithm with the PTAD framework, its real-time performance is improved by 13%, verifying the effectiveness of the proposed algorithm.

The limited number and high similarity of learning clauses lead to limited historical information and imbalanced search tree. In order to solve the problems, a dynamic branching strategy of award and punishment was proposed. Firstly, the variables of every unit propagation were punished. Different penalty functions were established according to whether the variable generated a conflict and the conflict interval. Then, in the learning phase, the positive variables for the conflict were found according to the learning clauses, and their activities were nonlinearly increased. Finally, the variable with the maximum activity was chosen as the new branching variable. On the basis of the glucose3.0 algorithm, an improved dynamic algorithm of award and punishment named Award and Punishment 7 (AP7) was completed. The experimental results show that, compared with the glucose3.0 algorithm, the rate of cutting branches of AP7 algorithm is improved by 14.2%-29.3%, and that of a few examples is improved up to 51%. The running time of the improved AP7 algorithm is shortened more than 7% compared with the glucose3.0 algorithm. The branching strategy can efficiently reduce the size of search tree, make the search tree more balanced and reduce the computation time.