With the development of Connected and Automated Vehicles (CAVs), related technologies are applied in many specific traffic scenarios. Especially in tunnels of highway, how to avoid traffic accidents attracts the attention of researchers. Therefore, this paper designs a novel multi-source sensing method for unmanned detection vehicle. Firstly, the Hector SLAM matching algorithm for multi-source sensing system is established by the fused data of camera, laser point cloud and inertial measurement unit (IMU), which can map the tunnel scenario and locate the real-time position. Secondly, a route planning algorithm in tunnels based on the combination of A* algorithm and dynamic window approach (DWA) is proposed. Thirdly, the detection method of cracks and fire in the tunnel is realized based on the improved faster region-based convolutional neural network (Faster R-CNN). Finally, through the field test in tunnel scenarios, experiment results have shown that the tracking and positioning error detected by the proposed method are no more than 0.2 meters. Meantime, the detection accuracy rate is 94.53%, and the recall rate is 86.33%, which can satisfy the detection requirement of driving safety in tunnels. It has provided an effective and feasible detection technology in tunnels based on autonomous vehicles.