<p>Vehicle lane change behavior recognition has become a key research area with notable achievements. However, existing methods often overlook spatial relationships between vehicles, resulting in poor recognition when multiple vehicles are present in the video. To address this issue, this paper proposes a novel vehicle lane change behavior recognition method, namely Multi-Scale Three-Stream 3D ResNets (MST3D), which effectively represents spatial information by depth maps during lane changes, thereby capturing the spatial relationships between vehicles on the road. Meanwhile, MST3D captures appearance information through RGB images and motion information using optical flow. By utilizing multi-scale image extraction and integrating three information streams, this method provides a more comprehensive representation of lane change behavior. 3D ResNets effectively capture spatio-temporal features, improving recognition accuracy in complex traffic environments. In addition, a data distillation method based on image aspect ratio and size is proposed to eliminate noise generated by distant and incomplete vehicles. On the Prevention dataset, the proposed method achieved a recognition accuracy of 90.94%, significantly higher than the accuracy of single feature recognition. Experimental results show that this method combines the advantages of each information stream and effectively improves the accuracy of vehicle lane change behavior recognition.</p>