*Article* **Electric Vehicle Fire Trace Recognition Based on Multi-Task Semantic Segmentation**

**Jiankun Pu and Wei Zhang \***

School of Microelectronics, Tianjin University, Tianjin 300072, China; 3015204014@tju.edu.cn **\*** Correspondence: tjuzhangwei@tju.edu.cn

**Abstract:** Conflagration is the major safety issue of electric vehicles (EVs). Due to their well-kept appearance and structure, which demonstrate salient visual changes after combustion, EV bodies are recognized as an important basis for on-spot inspection of burnt EVs and make application using semantic segmentation possible. The combination of deep learning-based semantic segmentation and recognition of visual traces of burnt EVs would provide preliminary analytical results of fire spread trends and output status descriptions of burnt EVs for further investigation. In this paper, a dataset of image traces of burnt EVs was built, and a two-branch network structure that splits the whole task into two sub-tasks separately concentrated on foreground extraction and severity segmentation is proposed. The proposed network is trained on the dataset via the transfer learning method and is tested using 5-fold cross validation. The foreground extraction branch achieved a mean intersection over union (mIoU) of 95.16% in the burnt EV foreground extraction task, and the burnt severity branch achieved a mIoU of 66.96% for the severity segmentation task. By jointly training two branches and applying a foreground mask to 3-class severity output, the mIoU was improved to 68.92%.

**Keywords:** deep learning; semantic segmentation; electric vehicle fire
