In recent years, the demand for efficient and accurate object localization technology has been increasing with the rapid development of computer vision, radar detection, and biomedicine. Traditional object localization technology mainly depends on image acquisition, typically adopting the ‘image first, localization later’ method, combined with object localization algorithms for positioning. However, image-dependent object localization is often limited by the resolution of image reconstruction. Due to the complexity of the object, scene, and environmental uncertainties, existing object localization technology faces many challenges, such as severe occlusion, blurring, changes in illumination, and distance variations, which may lead to object loss. Additionally, geometric deformation and attitude changes may result in positioning failures, thus reducing localization accuracy. In contrast, single-pixel detection technology shows unique advantages in object localization, especially in low-light detection environments, with strong robustness and anti-interference capabilities. Furthermore, most existing research on single-pixel, image-free detection in complex scenes focuses on known object information, which improves object localization performance in complex environments through algorithmic enhancements. However, there are relatively few studies on image-free target localization for unknown object information. Therefore, we aim to propose a single-pixel, image-free object localization method that works with unknown scene information. This method can effectively protect image privacy, save storage space, and provide a new approach for object localization.

We propose a single-pixel, image-free object localization strategy based on global search (SPIF-GSOL) for single-pixel imaging. We use an improved genetic algorithm to perform a global search when the object information is unknown. The template projection position is updated through continuous feedback iterations, and the optimized reference point coordinates are obtained. The template and the pure white pattern are projected onto the coordinate point, and the R-value is calculated at this stage. This process is repeated to obtain a continuously optimized population, and the projection position is updated iteratively. The method checks whether the number of iterations meets the termination condition. If the termination condition is satisfied, the highest R-value and the corresponding coordinate point position after convergence are output. If the termination condition is not satisfied, the global search is performed based on the fitness value, and the matching position is continuously updated until the iteration count is met, which ultimately achieves image-free object localization.

1) The method proposed in this paper can still achieve accurate positioning in the case of complex scene interference. As shown in Figs. 9?11 and Fig. 13, the method can accurately locate the object when the template position changes, its size changes, its attitude changes, or there is occlusion. 2) The proposed method (SPIF-GSOL) is compared with traditional ghost imaging (GI) localization, differential ghost imaging (DGI) localization, normalized ghost imaging (NGI) localization, and compressed sensing ghost imaging (CSGI) localization. Fig. 13(a) presents the time consumption curves for different projection times, while Fig. 13(b) shows the localization accuracy curves for the same projection times. The results reveal that the proposed method consumes the least time and achieves the highest accuracy compared to the “image first-localization later” method. 3) Compared to 74529 iterations without using the genetic algorithm, the number of iterations for our method is reduced to 5000, which represents a 93.29% reduction. This fully demonstrates that the single-pixel image-free object localization using this method effectively reduces the number of projections and avoids the problem of increasing projections with the growth of image size. 4) Compared with different fitness functions, the proposed method achieves the lowest mean absolute error (MAE).

In our study, we do not need to load a fixed projection sequence. We calculate the fitness value based on the received single-pixel value and use an improved genetic algorithm to continuously update the template pattern projection position during the feedback iteration process. Simulated and experimental results show that, when the object scene information is unknown, the proposed method does not require imaging to accurately localize the object in complex scenes, while also effectively protecting image privacy. The genetic algorithm also reduces the number of projections and improves the efficiency of object localization. Furthermore, the proposed method exhibits certain anti-interference characteristics for object scaling, rotation, and offset in complex scenes. Compared to other algorithms, our method achieves higher localization accuracy.