Abstract: Aiming at the problems of multiple interferences, similar false detections and dynamic drift in complex urban rail transit construction site environments with dust and low illumination, this paper proposed a multi-type engineering vehicle detection algorithm based on improved YOLOv8. The algorithm constructed a dual-modal feature extraction system in the Backbone network, extracted features through visible and infrared backbone networks, aligned features via a spatial transformation network, and realized weighted feature fusion through dual-channel attention. In the Neck module, it introduced deformable convolutions and a key component detection head. The network predicted component coordinates through an independent regression branch and adopted L1 loss and Generalized Intersection over Union (GIoU) loss for optimization. In the Head module, it used a Long Short-Term Memory (LSTM) network to perform temporal encoding on continuous video frames, and utilized Spatio-Temporal IoU (ST-IoU) to quantify temporal consistency and constrain detection continuity. Experimental results verified that the precision rate of engineering vehicle recognition increased from 91.822% to 98.625%, the recall rate rose from 74.777% to 86.073%, and the F1-score improved from 82.121% to 91.879%. This study provides a novel method for the monitoring of multi-category engineering vehicles in complex construction site environments.