Lightweight Vehicle Detection Model Based on Feature Map Similarity Measurement

With the development of deep learning technology, target detection has also made significant progress in transportation. Still, the computational complexity and resource consumption of the models used for target detection have become the main bottlenecks for practical deployment. To solve this problem, this paper proposes a model pruning method based on a feature map similarity metric. The method employs the YOLOv5s model and analyzes the similarity between filter feature maps to quantify the redundancy of filters and implement structured pruning. Pruning experiments were conducted on the UA-DETRAC and BIT-Vehicle datasets. The pruned model maintains a high detection accuracy on the UADETRAC and BIT-Vehicle datasets, with an [email protected] (mean Average Precision) of 97.83 % and 99.19 %, respectively. Meanwhile, the computational cost is reduced by 28 %, and the number of parameters is reduced by 30 %. Compared with the benchmark model, our method performs better regarding efficiency-accuracy balance and provides an effective solution for edge deployment of vehicle detection models.