A vehicle detection method based on cross-scale feature fusion

Abstract

This paper presents improved algorithms developed to enhance the detection performance of multi-scale vehicles across various lanes. The proposed methods specifically address the challenge that small-scale vehicles in video surveillance systems are susceptible to false positives and missed detections under complex conditions. These limitations ultimately lead to imbalanced detection outcomes across vehicles of different scales. Firstly, the cross-scale feature fusion structure is proposed to enhance the deep fusion capability of multi-scale features. This effectively addresses the issue of deep feature maps struggling to capture small-scale vehicle information caused by excessive downsampling. Secondly, the cross-scale feature fusion module is proposed to enable the model to dynamically capture features from multiple dimensions, facilitating a deeper understanding of both coarse- and fine-grained data and thereby significantly enhancing the performance of multi-scale vehicle detection. Additionally, the downsampling convolution is optimized using Receptive-Field Attention to improve the model's ability to understand the detailed features of multi-scale vehicles. Finally, the Wise-Intersection over Union (Wise-IoU) loss function is utilized to improve the detection performance for low-quality vehicle samples. Experiments on the VisDrone and Vehicle datasets show the number of parameters and model size of the proposed algorithm in this paper have been significantly streamlined. The algorithm can effectively balance the detection performance of multi-scale vehicles, thus obtaining higher overall detection accuracy. In addition, the test results on the DOTAv1 dataset show the proposed method has good generalization ability and cross-scene detection capability.