Small Moving Target Detection In Optical Image Based On Depth Optical Flow Estimation

Small moving target detection is widely used in agriculture,transportation,military and other fields.However,the dynamic background of many image sequences and the small size of the target in the practice increase the difficulty of moving target detection.Currently,there are some problems in the moving target detection algorithm,which are as follows: unsatisfactory effect of image registration in complex background,information loss of small targets due to the multiple down sampling of the optical flow estimation algorithm,and the long term of optical flow estimation algorithm.To address the above problems,an optical flow estimation algorithm based on local matching was proposed in this paper.Then based on this algorithm,Center Net was used to detect the optical flow anomaly area of the obtained optical flow field to achieve small moving target detection.Finally,this algorithm was applied in the infrared tail flame scene,achieving good detection performance.The research contents of this paper are as follows:(1)An optical flow estimation algorithm based on local matching was proposed to reduce the time consumption of image optical flow estimation.The method of feature matrix of two frames from feature extraction of two original images of the input optical flow estimation network introduced the target motion information and reduced the space scope of feature matching,which allowed any feature vector to perform feature matching only in its relevant local area,thus reducing the amount of data to be processed;A block based local matching strategy was also designed with and batch processing mechanism introduced to avoid the long term of processing data caused by point by point local matching strategy,so as to accelerate the algorithm under the premise of ensuring the quality of optical flow estimation.(2)A detection algorithm for optical flow anomaly area based on Center Net was designed to improve the accuracy and recall rate of the target detection.A normalization method of optical flow field was designed to magnify the optical flow difference between the optical flow anomaly area and the background where the moving target is located;The method of dimension expansion of optical flow field was designed to expand the 2D optical flow field to the 3D optical flow diagram,so as to keep consistent with the subsequent interface of Center Net;The optical flow diagram without moving target was constructed and added with the training set of Center Net to solve the problem that it is difficult to distinguish the amplified noise from the target due to normalization.(3)In the infrared tail flame detection scene with middle and long distances,the tail flame target could be detected with good performance by the algorithm in this paper after the simulation data set was processed appropriately.