Laser alignment is a prerequisite for stable energy acquisition at the receiver end in laser wireless power transmission systems. Laser power transfer imposes high requirements on alignment accuracy, stability, and real-time performance. Therefore, a laser alignment system design method is proposed, and optimizations are made to the region of interest extraction and image preprocessing methods. On one hand, the SSD (single shot multi-Box detector) model is improved by introducing MobileNet, incorporating spatial attention mechanism, and fusing semantics. The improved model is used for training and predicting the regions of interest. Compared to the original model, the training speed is improved by 71.67%, the model size is reduced by 52.48%, the real-time detection speed is increased by 295.30%, and the detection error is significantly reduced. On the other hand, the weights of grayscale conversion are optimized, and an adaptive threshold selection using a histogram is implemented. The elliptical fitting method and centroid method are employed to detect the spot and beacon center, reducing the error in spot localization. Experimental results show that the improved laser alignment system achieves a stable accuracy of over 95% and meets the requirements of accuracy, speed, and stability in application.