A Study On Methods Of Target Identification And Tracking Based On Deep Learning

The development of computer vision is a basic work of intelligent development in the future.In order to integrate intelligent machines into human life,they need to observe the world like human beings so one of their basic capabilities is to identification and tracking.Now with the help of deep learning,the effect of tracking has achieved exciting results,but deep learning itself has a lot of unsolved mysteries,and the tracking based on it still has the potential.On the one hand,in deep learning,network training always takes too much time;on the other hand,the tracker with deep learning still can be improved in speed and precision.Therefore,this paper firstly studies how to reduce the training time of deep learning,and then studies changing an online deep tracker to an offline tracker.Finally,the outline information is used to improve the tracking accuracy of the deep tracker.The main work of this paper is as follows:Firstly,the impact of data distribution of small batches of data on network training was studied.The training time of the deep model is long,largely because the network loses up and down during the training,which causes a waste of time.Especially when using the random gradient descent method,the network is always easy to oscillate near the best point,wasting training time.In contrast,using the batch gradient descent method,although the network convergence is good,the hardware burden is large and the speed is slow.The existing solution is to use a compromised small batch gradient descent method.However,actually,using this compromise method,there will still be serious fluctuations in network losses.By designing experiments and comparing the experimental results,this paper explores that the data distribution of small batch training data has a great influence on the network training.The results show that if the network loss is more smoothly reduced,at least the adjacent small batch data should be made consistent in data distribution,and the best case is to be consistent with the overall training data,so that the training data can be fitted faster.Secondly,the off-line processing of the online deep tracker(MDNet)is studied.On the one hand,the deep learning-based tracker itself has no speed advantage because of the large number of parameters not to mention he complicated online update process.On the other hand,in general,the offline tracker is not as good as the online class tracker,but its tracking process is simple and speed is dominant.Therefore,studying how to turn a good online tracker into a good offline tracker is valuable.Experiments on the tracking test platform show that the final offline tracker is competitive.Last but not least,this paper proposes a deep tracker with outline response map.Many deep trackers use neural networks to extract relevant features of the target,and then use the feature as a matching template to perform target matching in the search region of the current frame and determine which parts of the current frame are foreground while the others are not according to the strength of the position response.Based on this,this paper visualizes the target features extracted by the neural network,and combines the features extracted by the traditional machine learning tasks to study the correlation between them then uses the outline information of the target to obtain the position response of the target in the search area via a method of filter.At last superimposing it with the position response obtained by the network,thereby achieving the purpose of suppressing position noise and improving the tracking accuracy.The test results on the target tracking benchmark platform and the visual results of the target location response show that the method is feasible.