TOSS: Deep Learning-Based Track Object Detection Using Smart Sensor

In high-speed railways, train collisions with obstructions on the trackside are prevented using automated railroad security systems. Rail safety is being improved, and accident rates are reduced through continuous research. The rapid advancement of deep learning (DL) has created new possibilities for research. In this article, a novel track object detection using smart sensor (TOSS) approach has been proposed for tracking the objects in railway track (RT) using DL networks. A TOSS approach uses a camera and light detection and ranging (LiDAR) as primary sensors for detecting objects and faults in RT to prevent accidents. Preprocessing methods include data cleaning, min-max normalization, and calibration to ensure data quality by removing unwanted data from datasets. Then, clustering the preprocessed data to determine objects that are initial sizes and positions. In visual data processing, the camera images are denoised using a bilateral filter (BF) to remove noise. In order to prevent accidents on the RT, the YOLOv8 network is utilized to accurately localize and detect objects on the track. The visual and digital data from the camera and LiDAR sensor are given as an input to the fuzzy system. This data will be used to generate the system alert message that is sent to the loco-pilot and nearby control rooms. In the experimental analysis, the proposed TOSS approach achieved an overall accuracy of 98.91% and an mean average precision (mAP) of 97.1% for detecting objects and faults efficiently. The TOSS approach demonstrates significant progress in the overall accuracy range by 13.86%, 10.22%, 5.46%, 8.8%, and 1.50% better than 2-D singular spectrum analysis (SSA) + Deep network, YOLOv8, YOLOv5s-VF, FR-CNN, and YOLO-GD, respectively.