Tp-yolov8: a lightweight and accurate model for traffic accident recognition

Zhaole Ning, Tianze Zhang, Xin Li, Aiying Wu, Gang Shi

Abstract: Traffic accident detection is an important part of road safety, impacting the lives of those involved and others on the road. Using surveillance cameras on traffic poles to detect accidents poses unique challenges, such as incomplete dataset categories, small-sized detection objects, and the need for lightweight models. Current traffic accident recognition algorithms, while effective in detection, often require extensive resources, making deployment on edge devices difficult. This paper proposes a more accurate and lightweight traffic accident recognition model based on YOLOv8, optimized for traffic pole monitoring and deployment on edge devices. To improve small object detection, we made improvements to the neck. We modified the neck by adding a detection layer for small-sized objects using large-scale feature maps, along with a dedicated small object detection head (SODL-SODH). Additionally, we design a lightweight cross-scale feature fusion module (LCSFFM) to optimize the PAN-FPN structure, reducing model parameters and computational complexity while enhancing small-target detection. In the downsampling layer, we incorporate the squeeze-excited aggregate spatial attention module (SEASAM) into the C2F module to help the network focus on essential image information, with minimal impact on model parameters and computational complexity. To address dataset limitations, we built the traffic accident-type (TAT) dataset for training and evaluation, and validated it against other advanced methods. Experimental results show that our model outperforms the baseline on the TAT dataset, improving the mAP0.5 by 1% and reducing parameters by 25.9%. On the BDD-IW dataset, our TP-YOLOv8s outperforms other methods in terms of accuracy. Compared with the best other methods, it improves the mAP0.5 index by 1.4% and reduces the number of parameters by 84.1%.