TC-YOLO: Enhanced Vehicle Detection Approach for Traffic Surveillance Cameras Based on YOLOv8

Sheng Wu, Zhihao Peng, Linquan Yang

Abstract: With the popularization of traffic cameras in cities, the use of traffic cameras for vehicle detection is very crucial for traffic management. However, the pixel resolution of images captured by traffic cameras is generally low, and there is a significant difference between the data collected during daytime and nighttime. Existing object detection algorithms don’t perform admirably on low-resolution images. To overcome these challenges, we present a new and advanced one-stage detector referred to as TC-YOLO (traffic camera-yolo). Our methodology integrates the Global Attention Mechanism (GAM) and Deformable Convolutional Modules (Dconv) into the backbone and head of YOLOv8. This integration significantly improves the extraction of crucial features. Additionally, we make alterations to the model arrangement through the introduction of an additional detection head. This modification enables feature aggregation across multiple scales, particularly benefiting the detection of small objects. We assessed the effectiveness of TC-YOLO on our traffic camera detection dataset, and the results obtained from the experiment substantiate the efficacy of our model. The model attained a mean performance mAP (mean Average Precision) of 0.80, yielding Precision and Recall rates of 0.91 and 0.93, correspondingly. The outcomes of this study outperform those of existing competitors in this field. The findings of this research constitute a substantial contribution to the advancement of traffic camera image data detection.