Research On Target Detection Based On Background Motion Compensation

Target detection technology under motion platforms plays a significant role in military and livelihood fields,and has high practical value and research significance.In motion platform scenarios such as handheld devices,airborne or vehicle-mounted facilities,dynamic background interference and sudden changes in ambient lighting occur,increasing the complexity of target detection and resulting in lower detection accuracy and operational efficiency.Existing motion target detection techniques are effective in static scenes,but most of them are not applicable to dynamic backgrounds.To address these problems,this thesis conducts research experiments on target detection techniques based on background motion compensation,and the specific work is as follows.（1）A background motion compensation technique based on inertial information and improved Oriented FAST and Rotated BRIEF（ORB）algorithm is designed and implemented.The external synchronous fusion of inertial information output from inertial sensors and video image data is designed to achieve the accurate matching and fusion of inertial data and image data.The initial compensation frame image is constructed using the platform motion parameter information from the synchronous output,and the original image is regionally filtered according to the results after differencing to reduce the area where the subsequent feature detection algorithm acts and reduce the operation time consuming,which improves the operation efficiency by about 23%.The ORB algorithm is improved by using the DAISY descriptor,which is insensitive to the rotation and light intensity changes between images,and optimizing its principal direction assignment method to replace the BRIEF descriptor,which improves the matching accuracy of the algorithm when the rotation and light intensity changes occur in the images The accuracy of the algorithm is improved by about 44%when the image is rotated and the light intensity changes.The K-Nearest Neighbors feature point matching algorithm is improved by combining the Euclidean and Barthian distances to measure the similarity of feature points,which further reduces the false matching rate of the algorithm.The six-degree-of-freedom motion information obtained from the inertial sensor is used to determine the affine motion model,and the optimal feature point pairs are selected according to the Progressive Sample Consensus algorithm.After solving the affine motion model,an accurate compensated frame image is constructed to accurately remove the background motion interference.It is proved that the compensation accuracy is higher and the operation time is lower compared with the traditional background motion compensation technique,and the overall performance is excellent.（2）A motion target detection algorithm based on the improved frame difference method is designed and implemented.Firstly,the two-frame difference method is improved,and the adaptive thresholding method is used in the difference to improve the separability of target and background in complex environments.On this basis,the target edge information is extracted by Canny edge detection algorithm,and the target information is enhanced by the cumulative difference method to reduce the hole phenomenon and improve the target integrity by logical"or"operation.Experiments prove that the improved frame difference method in this thesis results in a more complete external contour and more adequate internal information of the target,and the completeness is significantly improved.For false targets,the difference in motion between them and the real target is used to reject them by motion analysis.In order to reduce the computational effort,the optical flow calculation method is improved in the motion analysis,and the differential results are used to divide the region,and only the region where the differential response exceeds a certain threshold is calculated by the Gunnar Farneback optical flow method for dense optical flow,and the pseudo-motion targets are rejected according to the results to improve the target detection accuracy.（3）Experimental testing of the detection effect of this thesis’s target detection technology is carried out in four sets of video sequences taken under the motion platform,while comparing with the traditional algorithm to analyze and evaluate the experimental results from both subjective perception and objective quantitative perspectives.Compared with the traditional algorithm,the detection result of this thesis is more complete,the number of interference items is significantly reduced,and the accuracy is significantly improved.The F₁ value of this algorithm is increased by 20.01%,the Percentage of Wrong Classification value is reduced by 36.40%,and the average time spent in a single frame is reduced by 27.78%,so the detection results are better in different application scenarios.The significant improvement of the accuracy and real-time performance of the target detection and the obvious enhancement of the robustness of the algorithm in this thesis demonstrate the effectiveness of the motion target detection technique in this thesis.