智能科学与技术学报 (Dec 2023)

Vehicle detection and recognition algorithm based on function improvement of YOLOv3

  • SONG Huajie,
  • ZHOU Lei

Journal volume & issue
Vol. 5
pp. 535 – 542

Abstract

Read online

YOLOv3 algorithm is characterized by high detection accuracy and high speed in the aspect of target detection and recognition. It is outstanding among similar algorithms, but there are still obvious problems. The loss operation of the wide and high part of the loss function of this algorithm is not obvious to the distinction between the big target and the small target, which leads to the problem that the loss calculation is not accurate enough and the prediction box size of the small target is not accurate. In view of this defect, the loss function of YOLOv3 algorithm was improved, the wide-height coordinate error was modified into the form of proportion, the non-maximum suppression method of the original model was improved and the overlap threshold was changed from the fixed value to the form of attenuation function. Finally, the model was applied to vehicle detection. The experimental results showed that the improved model improved the accuracy rate, recall rate and F1 value to different degrees while basically not affecting the vehicle detection speed, making the overall performance of the model better than the original YOLOv3 model.

Keywords