Research On Improvement Of Small Moving Target Detection Algorithm Inspired By Insects In Natural Scenes

In recent years,with the rapid development of computer vision technology,the small target motion detection has become an important research direction of computer vision.Detecting small moving targets in natural environment is one of the most important abilities of many organisms,such as mate finding,hunting and other behaviors.Dragonflies with a relatively simple visual structure are one of the species with the highest hunting success rate,so the study of insect visual pathways is an important way to inspire the detection of small moving targets.In the research of insect vision system as a model,the classical insect-inspired target detection algorithm(Insect-Inspired Tracker,IIT)uses the "Elementary Small Target Motion Detector(ESTMD)" as its front-end to perform local motion detection in the video,and the back end gains the "prediction area" and finally marks the position of the maximum response value in the whole field as the small target position.With the advantages of simple model and strong real-time performance,the IIT model performs well in the process of small moving target detection in simple natural scenes,but the effect is poor in more complex scenes.In order to improve this bionic model,the following work is carried out in this paper:1.IIT model analysis.This paper deeply analyzes the modeling process of the IIT model,analyzes and verifies the adaptability of the front-end local motion detection stage of the IIT model,and finds that the single-parameter ESTMD model in the frontend local motion detection stage of the IIT model has weak adaptability to adapt to the changes.of different speeds and sizes of the target;The data association at the back end of the IIT model only focuses on one of the multiple possible trajectories of the moving target,and cannot focus on and gain multiple possible trajectories at the same time(in complex natural scenes,pseudo-target trajectories are often formed due to local disturbances),the performance of the IIT model for small moving target detection is tested in natural scenes,and it is found that the detection effect is not good in more complex scenes.2.Research on the improvement of the front-end local motion detection stage.After analyzing the internal relationship between ESTMD model,EMD model(Elementary Motion Detectors,EMD)and Gabor filter,it is found that the essence of the ESTMD model and the spatio-temporal Gabor filter are the matching of the changes of the spatio-temporal brightness pattern,and have inherent similarity.Inspired by the application of Gabor filter in Hmax model,an improved scheme for the local motion detection stage in the front end of the IIT model is proposed:manually set multiparameter ESTMD model.At the same time,the tuning properties of the singleparameter ESTMD model are analyzed,and the multi-parameter ESTMD model is analyzed and verified,which shows the effectiveness of the improved scheme of the multi-parameter ESTMD model.Inspired by the application of Gabor filter in convolutional neural network,a second improved scheme is proposed,a multiparameter spatiotemporal Gabor filter is obtained by training a three-dimensional convolutional neural network instead of the single-parameter ESTMD model.3.Research on the improvement of the back-end data association stage.Aiming at the problem that the prediction,gain and maximum method of the data association stage at the back end of the IIT model are easy to fail,the "K shortest path" algorithm is used to perform data association on the local motion detection results to complete the target detection task.The validity of the K-shortest path scheme is verified under simulation conditions,and the combination of the single-parameter ESTMD model in the IIT model and the K-shortest path scheme is used to detect small moving targets in natural scenes,which further shows that the K-shortest path data association scheme can be completed.ESTMD model output array for target detection task.4.Research on target detection algorithm based on K shortest path.The small moving target detection algorithms in natural scenes are constructed by combining single-parameter ESTMD model,multi-parameter ESTMD model,multi-parameter spatiotemporal Gabor filter and K-shortest path scheme respectively.Carry out simulation research on two K-shortest path-based target detection algorithms in natural scenes,the results show that the K-shortest path-based target detection algorithm has better performance than the IIT model when detecting small moving targets in complex scenes.The target detection algorithm of the multi-parameter ESTMD model and the K shortest path scheme is robust to changes in target size and speed,the K-shortest path data association scheme improves the adaptability of the target detection algorithm based on the ESTMD model to complex scenes.The multi-parameter spatiotemporal Gabor filter and K shortest path scheme are more accurate than the IIT model detection results.Finally,this research realizes the improvement of the classical small moving target detection algorithm inspired by insects.