Aerial Video Object Detection Based On Compressed Sensing Using Milti-objective Optimization

With the development of Computer Science and Technology, Automatic Controlling Technique and Robot Intelligence, all kinds of robots were used in various fields of human life. Miniature unmanned aerial vehicles(UAV) is one of them. UAV applications are more and more rich now, Aerial photography as the “eye” of the UAV, is a vital part in these applications. The process quality of aerial photography largely determines UAVs’ ability. Object tracking and detection in aerial videos is the foundation of aerial video processing. However, Object tracking and detection in aerial videos is still an open question due to the complexity of aerial videos.In this dissertation, we make research on object tracking of aerial video. We designed and developed the Aerial Video Object Tracking System, specific to the particular problems in aerial video, such as the target is small in aerial video, long distance between target and UAV, fluttering and abrupt changing of the picture because of the instability of UAVs, as well as plenty of occlusion and big change of background. The main research contents of this dissertation are as follows:First, a discriminative tracking algorithm based on compressed sensing is designed. In our algorithm, we get the sparse representation of candidates on the train set, then we compute the confidence value by the reconstruction errors of sparse coefficient vector on both positive set and negative set, which is used to measure the probability that which candidates is the target. The bigger the confidence is, the corresponding candidate is more likely to be the target.Secondly, generative model based on compressed sensing is designed for our system. We construct the dictionary using k-means clustering algorithm. The local future is obtained by the sparse representation of patches from candidates on the dictionary. The histogram is composed of sparse coefficients vector of each patch. We get the similarity for each candidate by using the histogram intersection function, which is also used to measure the probability of the candidates is the positive sample.Thirdly, we designed an algorithm to solve the L1 minimization in our system based on Accelerated Proximal Gradient(APG). And we implement it with OpenCV, with this algorithm, we can solve the L1 minimization effectively, and the performance of the system is improved.Fourthly, a collaborative model based on Multi-objective Optimization(MOO) is proposed to integrate the discriminative model and the generative model. In this dissertation, we formulate the collaborative model by taking both confidence and similarity into account based on multi-objective optimization(MOO). In this way, advantages of the two method were used. The good candidates will not be abandoned becauce of the collaborative model based on MOO, and the accuracy and robustness is increased in our system.Fifthly, the aerial video object traking system is implemented with OpenCV. In our system, four method can be used to track an object in aerial videoes include SDC, SGM, MOO collaborative tracking and fast collaborative tracking. In this system, we used affine transformation to get standardized pathes, k-means clustering algorithm to construct the dictionary for SGM, and we solve the MOO model by NSGA-II.Finally, an overall test is made to evaluate the aerial video object tracking system. We use 6 aerial videoes provided by DARPA VIVID and 2 aerial videoes taken by our own UAV in the canpus to test the tracking performance of our system. We analyse and compare the advatage and disadvantage of the 4 method provided by the system according to the test result. Meanwhile, we made comparison between our system and two state of art algorithm including TLD and L1 APG. The test result shows that our system has a good performance on accuracy and robustness in aerial video object traking.