Addressing the issue of low vehicle recognition accuracy stemming from dense vehicle targets and occlusions in complex traffic scenarios, a vehicle recognition algorithm utilizing highorder spatial feature aggregation is proposed. Initially, during the downsampling phase of feature extraction, the HSIDM module is devised to facilitate deeper feature aggregation and minimize the loss of fine details. Subsequently, within the feature fusion component, the DMFAM module is introduced to dynamically adjust the weights of features across various scales, thereby acquiring multi-scale contextual information and bolstering the model′s adaptability to diverse features. Following this, a decoupled REL-Head detector is formulated to disentangle classification and regression tasks, preventing task mixing and enhancing the learning capability and interference resistance of local features. Ultimately, the model presented in this paper is deployed on edge devices for testing. Experimental outcomes reveal that on the complex traffic scene datasets BIT-Vehicle and UA-DETRAC, the mean average precision (mAP) of our algorithm has improved by 0.7% and 3.9% respectively compared to the original model. Additionally, it operates seamlessly on edge devices, demonstrating effective recognition capabilities. This indicates that the proposed approach can effectively enhance the precision of vehicle identification and is suitable for use on constrained devices.