The vertical target imaging test is an important part of the dynamic performance assessment of armored vehicles, and the accuracy of its center positioning is crucial to the testing and finalization of weapons and equipment. In the standing target imaging test of armored vehicles, range testers often face a series of complex and thorny problems: 1) Environmental factors such as lighting conditions and weather conditions have a significant influence on the cross symmetry of the target; 2) The target image has low contrast. 3) The target cross feature is affected by the built-in markings, occlusions, and target damage of the imaging system, resulting in the loss of key feature information. The difficulty of fully controlling these factors reduces the accuracy of target center positioning in weapon effectiveness evaluation. This problem has long troubled range testers. Although the military target detection method based on deep learning can highlight the characteristics of military targets to some extent, the contradiction between high-precision fitting and generalization makes it not the optimal solution to solve the precise positioning problem. Most of the existing traditional methods only focus on the detection of the general area where the target is located, and there are no reports on the precise positioning of the center of degraded targets with asymmetric cross features, low contrast, and missing feature information. To solve this pain point, there is an urgent need to study a central positioning method for degradation targets.

We propose a high-precision detection method for degraded target centers based on self-attention envelope tracking. The method consists of two stages. In the coarse positioning stage, the target area is detected through the YOLO series network. In the fine positioning stage, first, the image enhancement algorithm SINE based on the sine function is designed to widen the grayscale spacing of the image and enhance the outline of the target center area information. Second, using the orthogonal symmetry of the target cross feature, we propose a multi-directional attention fusion algorithm. It can better capture and utilize the rich semantic information of the target area by performing self-attention matching on the target image and its own transposed matrix. It also integrates multi-directional semantic information to enhance the target cross features while highlighting the center area to complete the reconstruction of the attention area. Then the Hilbert transform is used to construct the analytical signal and obtain the envelope of the image, highlighting important features with clear directionality such as edges and textures. Envelope tracking is achieved by eliminating the conjugate antisymmetric part in the frequency domain, so that the symmetrical structure of pixels around the target center is more noticeable. Finally, the obtained attention weight matrix is mapped to the enhanced target image to generate a target feature image, and the center of mass is calculated through the first-order moment to complete the target center positioning. The method achieves real-time performance while ensuring accuracy.

The effectiveness of the proposed method in locating the center of the degraded target is verified through simulation experiments and field experiments. Given the imaging characteristics of degraded target images and considering the vertical target imaging test environment, the simulation data produced in cross-feature asymmetry, dark light environment, and occlusion are compared with the accuracy of positioning the cross center using the traditional centroid method. The simulation results show that in the case of asymmetric cross center, the root mean square error (RMSE) of the centroid method is larger and fluctuates significantly throughout the sequence; the RMSE of our method is smaller and more stable under different contrast conditions [Fig. 5(a)]. In a dark light environment, when the target contrast of our method is 0.75, the center-positioning RMSE accuracy is 0.04 pixel [Fig. 6(a)]. The centroid method is affected by low contrast, resulting in a decrease in positioning accuracy, but the RMSE of our method remains small, showing higher accuracy and robustness [Fig. 6(b)]. When the occlusion degree is 24.87%, the accuracy of our method is 0.15 pixel [Fig. 7(a)], which can effectively deal with the problem of missing target feature information caused by target occlusion, the RMSE in the entire sequence is smaller, and the accuracy is more stable. Field tests show that compared with the measurement method combining template matching and the centroid method, our proposed method can more accurately locate the target center under full-spectrum conditions, fully demonstrating the detection capability of the method under degraded conditions (Figs. 8 and 9). In the actual application of the shooting range identification field, the pixels are converted into angle-measurement accuracy by bringing into the system angular-measurement resolution, which is used to inspect the performance parameters of the shooting range equipment.

Aiming at the problem that the cross feature of the target image in the vertical target imaging test scenario is asymmetric, low contrast, and missing feature information, which causes image degradation and affects the accuracy of center positioning, we propose a high-precision detection method for the center of the degraded target based on self-attention envelope tracking. Through a large number of experimental analyses, the following conclusions are obtained: 1) The multi-directional attention fusion algorithm we designed makes full use of the orthogonal symmetry of the cross feature of the target and effectively solves the problem of less target feature information under degraded conditions. 2) Envelope tracking is introduced to construct analytical signals through the Hilbert transform, analyze the amplitude of grayscale changes and local intensity changes in the image, and enhance the accurate characterization of the target center feature, which not only makes the center positioning result more accurate and reliable but also improves the anti-interference ability of the method. 3) The method has high detection accuracy and speed. When the cross feature is asymmetric, the center positioning RMSE accuracy is 0.06 pixel; when the contrast is 3.17, the accuracy is 0.01 pixel; when the occlusion degree is 24.87%, the accuracy is 0.15 pixel; the detection speed reaches 270 frame/s, which can provide strong support for the neutral target imaging test of equipment dynamic performance assessment.