JCSE, vol. 18, no. 2, pp.69-79, 2024

DOI: http://dx.doi.org/10.5626/JCSE.2024.18.2.69

Detection Algorithm of Tram Track Obstacles Based on Improved-SSD

Yunming Wang, Yiang Zhou, Xianwu Chu, and Guodu Peng
School of Automation and Electrical Engineering, Dalian Jiaotong University, Dalian, China

Abstract: The accurate and rapid identification of tram track obstacles is a crucial aspect in improving the safety of urban tram driving. To improve the detection accuracy and detection speed of urban tram track obstacles, the current study proposes an urban tram track obstacle detection algorithm based on Improved-SSD. To this end, for Conv3_3, Conv4_3, and Conv5_3, a bidirectional fusion module is designed to strengthen the feature expression ability of the low-level feature layer and enrich the semantic information. Meanwhile, for Fc7, Conv6_2, Conv7_2, Conv8_2, and Conv9_2, a two-stage deconvolution module is devised to compensate for the lack of detailed information of the high-level feature layer. To improve the detection speed, the convolution split structure is designed to replace all 3횞3 convolutions in the backbone network VGG16. Then, to improve the model?셲 ability to match a specific dataset, the k-means algorithm is used to optimize the aspect ratio of the prior bounding box. Finally, the improved algorithm is trained with and tested using a selfmade dataset. The experimental results show that, compared to the traditional SSD, the mean average precision of the Improved-SSD algorithm in detecting track obstacles is increased by 1.09%. The detection speed is also increased by 0.9 FPS. Lastly, the prediction box matches the real obstacle box better than that of the traditional SSD.

Keyword: Tram; Obstacle detection; Deep learning; SSD; Prior bounding box

Full Paper:   165 Downloads, 424 View