Design Of Fast Directional Detection Algorithm For Small Vehicles From Aerial Images

Real-time monitoring of vehicle targets from a high-altitude overhead perspective is significant for intelligent transportation or security surveillance.Vehicle targets in aerial images are not well detected due to other factors such as the small proportion of images they occupy.In addition,the directional information of the vehicles from a high-altitude overhead perspective is very important,and it cannot be ignored for applications in specific situations。 However,direction assessment is a serious shortcoming for the existing target detection framework.This paper conducts real-time detection research on small vehicle targets from aerial imagery and proposes a series of solutions according to specific requirements.The final good evaluation results were obtained.The main research content of this article is as follows:1.The convolutional network with multi-parameter exponential activation is used to extract features.The target detection algorithm based on deep learning to achieve good detection results depends on the fine features extracted by the feature extraction network,so the quality of the feature directly determines the detection effect.In this paper,the active layer with the greatest influence on convolution filtering is the object of interest,and a multi-parameter exponential function is proposed as a new activation function.This function can better fuse features and can be freely converted between the rectification function and the exponential function.More importantly,its parameters can be updated,that is,the activation layer is incorporated into a parameter update system similar to the convolutional layer.This structure enables the network to better perform feature extraction.Multiple open source networks and public data sets were used for experimentation.The result is a good proof for the actual effect of the function.2.Real-time detector for small vehicles from aerial imagery.The improved feature extraction network is beneficial for extracting the subtle features of small objects from aerial images.This paper analyzes the main defects of the universal detection framework for small target detection,and proposes a lightweight scale fair single-convolution detector dedicated to vehicle detection in aerial imagery.The architecture is based on a single regression network and uses a feature pyramid model to predict frames.The form ensures that these detection frames can be matched to vehicle targets of different sizes and proportions.In addition,the proposed architecture has a smaller model and fewer layers than other common single network inspection architectures.Second,the fair scale allocation strategy is proposed.The strategy includes optimization of the sources of the prediction boxes and the prediction box size and scale.It is more suitable for the specific target.Finally,due to the small prediction frame caused by the fairness scale,the negative sample is greatly increased,and the maximum background score strategy is proposed for this purpose.Negative samples were scored and the highest scores were retained as negative samples.It is paired with difficult mining techniques to balance the number of positive and negative samples.3.Regional orientation of vehicles from aerial images.In the actual scenario,the direction assessment of the vehicle targets is very extensive for applications such as surveillance.Based on the lightweight scale fair single convolution detector,the angle information is increased to form the orientation lightweight scale fair single convolution detector.It uses the orientation border with angle information as the target data processing method,and rejects the common horizontal border.The orientation border can effectively reduce background interference in complex backgrounds and enhance the robustness of the detector.Finally,the experimental results show that the orientation detector can not only estimate the orientation of the vehicle targets,but also improve the detection accuracy.