Research On Small-scale Target Detection Method Of Aerial Video Movement For Reinforcement Learning

The wide application of aerial drones in various fields has brought convenience to people.The emerging industries with aerial drones as the core are rapidly expanding.At the same time,the research on visual technology in aerial video has been paid more and more attention by scholars at home and abroad.In the aerial video taken by aerial drones,the moving target units are usually tightly connected and the target is small.At present,the target detection method still has great challenges in this application scenario,due to the rapid movement of the drone itself and the flight process.The effect of jitter noise is often difficult to detect an object of interest in an image through a single frame of image.This paper mainly implements the algorithm improvement and optimization of the defogging and moving small target detection of aerial video on the UAV ground station platform.The main work and contributions of this paper:1.this paper proposes a dark channel fast defogging algorithm.In the calculation process of the dark channel transmittance map of the aerial image with fog,the amount of defogging operation is greatly reduced by using the mean down sampling and the bilinear interpolation method.During the experimental test,the method proposed in this paper is 2-3 times more efficient than the traditional dark channel prior method.In the small target detection of aerial video motion,this paper proposes to dynamically select the appropriate interval step size and accumulated frame number by dynamic inter frame difference(ORB-DFD)method.The global motion estimation of the background is obtained by matching the adjacent frames by the global ORB feature,and the interval step is obtained according to the translation component,the rotation component and the scaling component of the background.At the same time,the energy accumulation value of the pixel is calculated and the pixel distribution is counted,and the appropriate accumulated frame number is obtained under the constraint condition.Finally,the moving target information is obtained by the inter frame difference and the morphological method.In this method,in the aerial video motion small target test experiment,the detection accuracy is improved,and the frame rate also meets the real-time requirement.2.This paper proposes an aerial video motion small target detection method with target sequence confidence.The video sequence correlation information is added by the backward back off method,and the long-term sequence confidence description and pixel gray-scale confidence of the moving target are introduced to improve the accuracy and robustness of the moving target detection.The screening of candidate targets is constrained by a fixed-size sliding window,and then the pixel gray value distribution ratio of the sliding window region and the boundary expansion region is used for determination,culling,and frame integration.The backward regression yields a long-term value expectation as a long-term sequence confidence,and the long-term sequence confidence description and pixel gray-scale confidence are used to evaluate the moving target.Through the target average confidence evaluation method,the performance of different sliding window sizes and discount factors is obtained.In the aerial video test of different scenes,the detection accuracy of this method is improved under the premise of ensuring real-time detection.3.In this paper,the nonlinear strategy network of deep inverse reinforcement learning method is applied to the decision of moving target position and size information.The model includes feature network layer and strategy network layer.The feature network layer extracts the candidate target region by the multi-layer convolution feature information,and associates the current frame candidate target with the previous n frame detection result as the policy network initial layer.The strategy is expressed by the weight value of the full connection layer,and the optimization of the feature network model and the exploration strategy of the strategy network layer are iteratively updated with the assistance of the expert trajectory,and the nonlinear fitting of the reward function and the learning process of the expert strategy are completed.Bayesian posterior probability based on value estimation method is easy to cause indirect strategy degradation.This paper transforms the extreme value problem of strategy optimization objective function into proxy function extremum problem.The parameter domain vector of the strategy is optimized by the confidence domain strategy optimization method,and the average KL divergence of the state space and the action space distribution is introduced as a constraint condition to ensure that the strategy monotonously does not decrease during the update iteration process.The experience pool training method of aerial video clips is divided into two stages:the strategy network assists the expert strategy through the expert trajectory,and the strategy network autonomous improvement strategy.In the application of the UAV ground station platform,this paper proposes a dark channel fast defogging method,a dynamic frame difference detection method,a target sequence confidence detection method and a deep inverse reinforcement learning detection method.