LiDAR System for Rail Detection
Author： Neuvition, IncRelease time：2021-07-14 09:21:02
The safe operation of rail traffic is inseparable from safe driving of rail trains and daily maintenance, safe operation of rail subgrade. The layout integrating LiDAR system with rail subgrade and rail train will play an important role in the highly automated operation of rail traffic in the future. This layout will detect smaller target obstacles, can prevent foreign objects along the railway from causing harm to the safe operation of trains, and comprehensively improve the efficiency of rail transportation and reduce safety risks.
The active obstacle detection LiDAR system for subway trains uses the camera + LiDAR system solution that is currently widely used in the field of automobile automatic driving, to achieve high-precision positioning and active obstacle avoidance functions. This LiDAR system will be independent of the train signal control system, and the new generation of safety protection equipment will further promote the process of unmanned subway trains. The currently widely-used solution combines a far-focus camera (mainly responsible for detection in the range of 50m-300m) + close-focus camera (mainly responsible for detection in the range of 1m-50m) to achieve forward coverage of 1m-300m. The high-definition point cloud data collected by three-dimensional LiDAR is used as auxiliary data. Through video-point cloud fusion, relying on the powerful neural network deep learning algorithm of the vehicle-mounted host computer, the LiDAR system can realize the accurate positioning of the train and the accurate identification and detection of various obstacles, with an effective detection distance of 8m-300m, the detected object of not less than 20cm*20cm, the false-negative rate <0.01%, and the false-positive rate is <0.01%.
***The false-positive rate means that the obstacle detection system gives N times of obstacle alarms, of which M times are not obstacle alarms, recorded as P,
The calculation formula is: P=M / N*100%
The false-negative rate means that the obstacle detection system should give N alarms, of which M times failed to report, recorded as S, and its calculation formula is: S=M / N*100%
Object detection algorithm – YOLO
In recent years, the target detection algorithm has made great breakthroughs. The more popular algorithms can be divided into two categories. One is the R-CNN (R-CNN, Fast R-CNN, Faster R-CNN) algorithm based on Region Proposal. They are two-stage and need to use the heuristics method (selective search) or CNN network (RPN) to generate Region Proposal, and then perform classification on Region Proposal, which is more accurate, but slower. The second type of algorithm is faster, but the accuracy is lower. The other is one-stage algorithms such as Yolo and SSD, which only use a CNN network to directly predict the categories and positions of different targets. In the case of processing a 1280*720 resolution picture on the same system based on the NVIDIA Jetson TX2 platform, using Faster-RCNN series algorithm detection speed is about 5 frames, while using YOLO algorithm can reach a detection speed of about 30 frames. Because the algorithm is used for real-time detection of obstacles in the forward direction of the train, the maximum speed of the train is usually 80km/h, which requires the data update frequency should be above 10fps. The sensing system needs to detect obstacles 250-300m ahead, and the detected objects are not smaller than 20cm*20cm. YOLO can meet the obstacle detection needs of rail applications in terms of detection speed and detection accuracy.
Track detection algorithm UFLD
Lane line detection is a basic computer vision problem with a wide range of applications (for example, ADAS and autonomous driving).
There are two mainstream methods for lane line detection. The traditional method is image processing and the other is the image segmentation method based on deep learning. However, the lane line detection algorithm currently has two major difficulties: 1. The lane line detection algorithm based on image segmentation is a pixel-by-pixel task, which requires a large amount of calculation and is usually not suitable for real-time scenes of autonomous driving; 2. for the positioning of the lane line, no-visual-clue can only be well positioned by relying on the global information of the surrounding traffic direction. At the same time, challenging scenes with severe occlusion and extreme lighting conditions correspond to another key problem of lane line detection.
Based on the above difficulties, the track detection algorithm UFLD proposes a new solution specific to extremely fast and no-visual-clue lane detection. The UFLD algorithm treats the lane detection process as a row-based selection problem using global features. Specifically, the formula is to use the global features to select the lane line position of the image on the predefined line, instead of segmenting each pixel of the lane line according to the local receptive field, that is, the lane line detection is defined as finding the position collection of some rows of the lane line in the image, that is, row-based classification, the schematic diagram of the selection is shown in Figure 2.
Since the algorithm solution is row-based selection, assuming that the selection is made on h rows, only the classification problems on h rows need to be processed, but the classification problem on each row is W-dimensional. Therefore, the original HxW classification problems are simplified to only h classification problems, and since the positioning on which rows can be set manually, the size of h can be set as needed, generally, h is much smaller than the image Height H.
In this way, the number of classifications is directly reduced from HxW to h, and h is much smaller than H, not to mention that h is much smaller than HxW. Therefore, the method mentioned in this article reduces the computational complexity to a very small range, solves the problem of slow segmentation speed, and greatly increases the lane line detection algorithm’s speed, which is also why it can reach 300+FPS.
Neuvition Titan M1-R
Neuvition Titan M1-R has 700 vertical lines and an effective detection distance of up to 600mm. Based on 1550nm laser echo signal measurement technology, it collects point cloud data in front of the train. The LiDAR system supports access to IMU acceleration data and GNSS high-precision GPS information. Combined with the point cloud stitching algorithm, it can generate a 3D high-precision railway line map, and generate a data map through calibration. Through the new point cloud data acquired every day sent to the system host, and the LiDAR system calculates, analyzes (compares with the data map), and stores the new point cloud data obtained by the LiDAR through the algorithm, it is mainly used for the detection of short-distance track lines, intrusion warning, and obstacle detection.
Neuvition LiDAR has the technical characteristics of high resolution, high detection accuracy, and accurate echo intensity. At the same time, it takes into account the angle coverage and angular resolution in the pitching direction, and achieves the following effects:
–Effectively resist the interference of ambient light intensity on detection;
–The vertical field of view takes into account coverage and grid resolution, and the maximum angular resolution is H0.01°*V0.01°;
–Industrialized vehicle design, has third-party certification test reports that comply with the motion, vibration, electromagnetic interference, temperature, and humidity environment of the train vehicle platform.
Detect the intrusion of foreign objects along the track
Fastener identification &Detection
Track fastener identification, check whether the fastener function is intact.