The advancement of electro-optical observation and targeting systems has introduced new challenges for electro-optical countermeasures. Laser-based active detection of cat-eye targets has become a prominent research focus in this field due to its capability for 24/7 monitoring of optical devices including sniper rifle scopes and binoculars, demonstrating considerable practical significance and academic value. However, the small size, significant scale variations, and lack of distinctive texture features of cat-eye targets present significant challenges for reliable detection and recognition, often leading to high false alarm rates in complex environments. To address these limitations, we propose a novel target detection algorithm specifically designed to suppress false alarms while ensuring robust detection performance.We propose a Decision-level Fusion based on Spatial Context (DFSC) algorithm for cat-eye target detection. The algorithm consists of three modules: In the cat-eye target detection module, an image binarization method based on adaptive iterative maximum inter-class variance is proposed. This method incorporates an iterative foreground refinement strategy and a dynamic convergence mechanism to accurately extract the connected regions of cat-eye targets while preserving local detail information. Additionally, feature descriptors based on the Fourier power spectrum and normalized weighted centroid deviation are constructed to enhance the discriminability between cat-eye targets and interfering objects. Furthermore, a multi-dimensional feature weighted fusion method based on adaptive environmental perception is introduced to achieve adaptive optimization of feature weights. In the general object detection module, deformable convolution (DCNv3) is integrated into the C2f module of the YOLOv8 backbone network, improving detection performance for occluded and small targets. In the decision-level fusion module based on spatial context, the spatial relationship between cat-eye targets and interfering objects is evaluated by calculating the occlusion rate of cat-eye targets, effectively suppressing false alarms.To evaluate the performance of the proposed algorithm, a comparative analysis is conducted against several representative algorithms in recent years, including three traditional methods and two deep learning-based methods. The evaluation is carried out quantitatively from five aspects: recall, miss rate, precision, false alarm rate, and algorithm runtime. The proposed DFSC algorithm effectively combines low-level visual features from target local regions with high-level spatial contextual information of interfering objects, significantly mitigating the impact of distractors in complex environments. Experiments conducted on a self-constructed cat-eye target detection dataset under complex environmental conditions show that, compared with existing mainstream algorithms, the recall rate is improved from 92.2% to 98.9%, and the precision is increased from 49.0% to 74.5%. The algorithm achieves a processing speed of 8.3 ms per frame while significantly reducing the false alarm rate in complex environments. Since the background environment in the field experiments is relatively simple and lacks interference from complex high-reflective targets such as pedestrians or vehicle lights, the algorithm is able to achieve stable detection of cat-eye targets at various distances, with virtually no false alarms. In dynamic scenarios, when a cat-eye target happens to be located at the edge of a driving vehicle and nearly overlaps with the position of the car's taillight, the algorithm may mistakenly classify the cat-eye target as an interfering object, leading to a drop in recall. However, this scenario represents a rare and atypical case, thus having a limited impact on the overall performance of the algorithm.To address the performance limitations of cat-eye target detection in complex environments, we propose the DFSC algorithm. The method integrates an adaptive iterative Otsu image binarization approach with a novel feature descriptor combining the Fourier power spectrum and normalized weighted centroid offset. An adaptive environment-aware multi-dimensional feature fusion strategy is introduced to dynamically optimize detection robustness. Furthermore, DCNv3 is incorporated into the YOLOv8 backbone network, leveraging spatial contextual relationships between interference targets and surroundings to enhance decision-level fusion. Experiments were conducted on a cat-eye target detection dataset built using a self-developed laser active detection system. The results show that, compared with existing mainstream algorithms, the proposed method improves the recall from 92.2% to 98.9% and the precision from 49.0% to 74.5%, with a per-frame processing time of 8.3 ms. The results demonstrate that the proposed algorithm significantly reduces false alarm rates while maintaining high computational efficiency, providing an effective and reliable solution for cat-eye detection in complex scenarios.