Abstract: To tackle the problems of low accuracy of detection and localization for traffic police object in complex
traffic scenes, a method to detect traffic police object based on the optimized YOLOv4 model was proposed in this
study. Firstly, four random transformation methods were employed to expand the self-built traffic police data set, so as
to solve the problem of model over-fitting and improve the generalization ability of the network model. Secondly, the
YOLOv4 backbone network was replaced with the lightweight MobileNet. The Inception-Resnet-v1 structure was
introduced to reduce the parameter numbers and deepen the network layers of the model effectively. Then, the
K-means++ clustering algorithm was adopted to perform clustering analysis on the self-built data set. In doing so, the
initial candidate frame of the network was redefined, and the learning efficiency was improved for traffic police object
depth features. Finally, to address the problem of the imbalance of positive and negative samples in the process of
network training, the focus loss function was introduced to optimize the classification loss function. Experimental
results demonstrate that the size of the optimized YOLOv4 model is only 50 M and the AP value reaches up to 98.01%.
compared with Faster R-CNN, YOLOv3, and the original YOLOv4 model, the optimized network has been
significantly improved. The proposed method can effectively solve the problems of missed detection, false detection, and low accuracy for traffic police object in current complex traffic scenes.

Key words: traffic police object detection, YOLOv4 model, K-means++clustering algorithm, deep feature learning,
focus loss function