Aiming at the problems of limited computational resources and low accuracy in long-distance small target detection in vehicle vision system, an improved lightweight network Ghostnetv2-SSD network is proposed to improve the computational efficiency and small target detection performance. Firstly, the VGG16 in SSD is replaced by Ghostnetv2 to reduce the model parameters and the computational volume, then the multi-scale feature fusion method is used to strengthen the recognition ability of small-size targets, and at the same time, the a prior bounding box is re-designed to make the model have a better detection effect; experiments through the \(nuScence\) dataset on CPU show that a five-fold increase in the test frame rate of the improved Ghostnetv2-SSD model over the traditional SSD model. The experimental results indicate that the Ghostnetv2-SSD model not only adapts to the limitations of vehicle hardware, but also improves the detection of small and medium-sized targets.