YoLite+: a lightweight multi-object detection approach in traffic scenarios

Abstract

The main challenges of traffic object detection are the density of individual traffic objects and their vulnerability to environmental factors such as weather and light intensity. For the baseline YOLOv4 network with large computation, it is difficult to run efficiently on GPU resource-constrained devices. Based on this question, we propose a lightweight traffic object detection approach, called YoLite+. In this approach, MobileNet and depthwise separable convolution are used to compress the parameters of the baseline YOLOv4 network. Then the approach utilizes the negative phase information of the feature map to improve the model’s accuracy, with almost no increase in the number of parameters. Finally, a data enhancement scheme for multi-scene simulation is proposed to expand the sample diversity and improve the model’s generalization ability. The experimental results show that our approach improves the inference speed by nearly three times over the baseline YOLOv4 and reduces the number of parameters by five times, while the accuracy meets the requirements of the business scenario.