Identifying occupied rail crossings has been an endeavor that has gone on for over 100 years. Cross bucks where invented in 1867 when roads where dirt and traffic was horses. At that time humans cross guards where often placed at crossings to direct horse and pedestrian traffic. The train horn was an effective advance warning system since people did not travel in vehicles with closed/sound deadening windows and Apple earbuds. In an effort to provide improved traffic management, sensors and communication systems have been piloted with varying degrees of success.

Physics of Motion

The motion of a vehicle like a train can be described by a basic equation of motion where the future location x depends on the current location x0, current velocity v0, current acceleration a0 and the time lapsed from this point.

Sensors can measure various aspects of the motion like current position (x0) or current velocity (v0) or current acceleration (a0). In general, unless the train is traveling at a constant velocity (a0=0) it is difficult to actually know the position of a train in the future. This is not a real train. Real trains vary the speed and direction continuously.

Characteristics of Rail Traffic

Freight trains consist of individual rail cars that must be built into trains sorted/arranged by each rail car’s destination. A rail yard is a complex series of railroad tracks for storing, sorting, or loading/unloading, railroad cars and/or locomotives. Different types of yards include main, shop, scale, cleaning, spur, siding, and storage.

Thru Freight Trains

Freight traffic that passes a crossings without stopping or turning around is called a thru train. Traffic between stations and yards is typically thru traffic. The average speed of a freight trains is 26 mph in 2021. The average length of a freight trains is 5040 ft. However, 1% of freight trains reach a length of 14,000 ft. Freight trains can reach speeds of 59 mph away from urban areas and switching yards. In yards, in cities, over complex curvature, and with very long trains the speeds can be 5 to 15 mph.

Switching Freight Trains

Switching trains are a characteristic of the process of building and sorting the cars on a train. These trains very in length, can stop on crossings, and typically change directions multiple times across an intersection.

Figure 1: Data from LinqThingz sensor illustrates stop-go nature of switching trains.
Figure 2: Switching traffic occupies the crossing as much as 20 minutes on a regular basis (measured with LinqThingz sensors).

Transit Trains

The primary type of Transit Train in the US is Amtrak. However, large cities like Chicago (Metra) and San Francisco (BART, CalTrain) have their own commuter train systems. Commuter trains are typically shorter and faster than freight trains. A typical Amtrak train is 85 ft and travel as speeds around 79 mph. The delay times at grade crossings is typically quite different as well.

Figure 3: Amtrak traffic blocks crossings only about 40 seconds in Milwaukee compared to freight trains with an average block time of 10 minutes and block times as long as an hour (measured with LinqThingz sensors) .

Types of freight cars

Trains contain multiple types of rail cars. The effectiveness of rail detection technology will vary with the type of car.

Figure 5: Different types of rail cars requires a wide range of detection parameters.

Presence verse Prediction

Historically, crossing safety devices operate primary on concept of Presence. Drivers will not generally have much information about when the blockage will occur or for how long. This lack of knowledge results in frustration for many drivers. Congestion related frustration leads to risky driving behavior.

A report to the US Senate …”In fact, in 2017, 71 percent of fatal crashes at public grade crossings occurred at those with protective gates. States often report no difference in crashes after these safety improvements were added, and some states have even reported a slight increase in crashes.” …from page 6 “Railroad Crossing Congestion and Its Impacts on Safety and Efficiency”, prepared by the US Senate Committee on Commerce, Science and Transportation, Chaired by Maria Cantwell, 2021.

A survey in Wisconsin Rapids illustrates that 94% said “YES” to the question “When approaching an intersection blocked by a train, have you turned around, used neighborhood streets or tried to beat the train to an unblocked at-grade crossing to avoid being delayed?”

Here is a video sent by a community member stuck in a mile-long traffic queue at an occupied crossing on Pilgrim Road in Brookfield, WI.  The video reflects the behavior reported in the survey.

Prediction is the advance warning about blocked crossings. With sufficient advance warning road traffic can be rerouted FAR IN ADVANCE OF APPROACHING THE CROSSING and thus allow the road traffic to avoid blocked crossing, the congestion and the safety risks associated with it.

For a fire fighting crew leaving the firehouse…it is not important to know that a crossing two minutes ahead is blocked NOW. It is imported to PREDICT if the crossing will be blocked when the firetruck reaches it two minutes from now.

Math vs Machine Learning

If trains always traveled the same speeds and always accelerated at the same rate. Then, it would would be a fairly easy task to estimate arrival time at a crossing based on the position at some previous time. However, train speed and acceleration very significantly based on rail condition, length of train, type of freight, other rail traffic etc. Machine learning is approach that determines speed and acceleration based on a plethora of external parameters. The parameters could include: current speed, current acceleration, which rail (for multiple rail configuration) which direction, what type of freight, previous station to station travel times, etc. Machine learning involves a computational process that takes in many types of historical input data, “learns” the dependency of result based on the input data, and creates a predicted results based on applying these dependencies on current parameters. The result can mean much more accurate “statistical” prediction that are more accurate than human derived formulas.

Sensor Descriptions

Here is a partial review of sensors used for this purpose.

Fixed-Distance Warning Time (FDWT) System

This system consists of a train detector placed a fixed distance away from the HRGC. The distance is long enough so that the fastest train is detected with a warning time of at least 20 seconds (Forsberg 2012; Halkias and Eck 1985). The 20 seconds warning time is the specified minimum standard in the Manual on Uniform Traffic Control Devices (MUTCD). Since the distance is set with the fastest train arriving at the HRGC with the minimum required warning time, oncoming trains at slower speeds may cause excessively long warnings at an HRGC. As the amount of warning time varies for each train having a different speed, drivers may make poor decisions to cross the HRGC (Cho and Rilett 2003).

Constant Warning Time (CWT) System

CWT improved on the FDWT system by taking into account train speed at the detector location. Similar to the FDWT system, a detector is placed from the HRGC at a distance that will provide a 20 second warning with the fastest train expected on the tracks. The system estimates the arrival time based on the measured speed along with distance from the crossing and provides a constant warning time for each train regardless of its approaching speed (Forsberg 2012; Halkias and Eck 1985). However, the system assumes constant train speed (i.e., no acceleration or deceleration).

Time Domain Reflectometry (TDR)

A more advanced system of the first generation technology was developed and tested by Turner (Turner 2009). The author developed a system using a concept called Time Domain Reflectometry (TDR) to detect oncoming trains and estimate their arrival times at HRGCs. The proposed method used rails as a two-wire differential transmission line. The system transmits electrical coded pulses through the railways at a known speed. When the train’s axles reflect these pulses, control systems detect the variation of the pulses. The distance from the control systems to the detected train axel is then identified based on the reflected time. A field test was conducted to ensure the feasibility of the system after properties of the electrical transmission line were determined based on variations for tie type, track ballast quality, and moisture content. Findings revealed that trains could be detected at theoretical distances of up to 5 miles under ideal conditions. However, poor conductance on railways was found when rail tracks were covered with dirt and mud. Moisture accumulation on rail tracks also significantly affected the detection range by deteriorating the quality of electric pulses transmitted.

Radar Detection Technology

Train detection systems using radar technology is the most popular second generation technique. A radar detection system uses electric signals beamed to surrounding areas at known propagation speed, and measures reflected time of the signals from a moving target object to identify information such as speed and direction. In analyzing low-cost active warning devices, a study was conducted using a radar detector (Roop et al. 2007). The study used Doppler radars located 0.5 miles away from the target HRGC to activate train-warning devices automatically. The radar system was installed for both directions so the departures and arrivals of trains could be detected in advance. The obtained data of the tested system was compared to the data from a track circuit-based device, which provides reliable detection. The result showed that even though the radar system successfully detected crossing trains at 100 percent accuracy, it produced many false positive detection, recognizing some non-train objects as the presence of a train. The use of radar at HRGCs can also be found in many other research papers (Chen 2015; Cho and Rilett 2003; Estes and Rilett 2000; Goolsby et al. 2003). These studies used Doppler radar in similar ways to collect train detection data at HRGCs.

Acoustic Detection Technology

Roop et al. (2007) conducted a study using an acoustic train detection system, which utilizes sound generated from train operation. The authors investigated frequencies of existing train horns and used frequency ranges to identify the presence of a train. The developed system was comprised of a power supply, an analog microphone, an analog amplifier, an analog-to-digital converter, a digital microprocessor, a logic controller, and a digital electrically programmable read-only memory (ROM). The equipped microphone was omnidirectional to detect trains at both directions by distinguishing the frequency ranges from peripheral noise. The authors conducted a field study to reveal the effectiveness of the acoustic detection system by comparing it with a track circuit-based detector. The result showed that the total activated number of the acoustic system was 26,094 during the field research period, 1,486 of which were correct detections of train presence. From the result, the authors indicated that the false alarm rate of the system was too high (94.3 percent) even though the system did not make any true negative detection. With the high degree of false positive detection rate, the authors concluded that advanced technical updates should be added to be applicable at HRGCs.

Magnetic Detection Technology

A research effort was made to examine the applicability of using Anisotropic Magneto-Resistive (AMR) magnetometers to determine when to activate warning devices at HRGCs (Brawner and Mueller 2006). The developed idea is based on the earth’s magnetic field, which is locally formed as parallel lines. When a large metallic object such as a vehicle or a train changes the regularly arrayed magnetic parallel lines, magnetometers detect the transformed traits of the lines. The proposed AMR sensors were not only able to detect train presence by recognizing the transformed array of earth’s magnetic field but also to differentiate the unique magnetic signature of each moving object. This unique signature of a metallic object can also be used to determine the moving direction since the wavelength is a mirrored image, as shown in Figure 2.3. The authors concluded that the 3-axis measurement capability of the AMR magnetometers is feasible to detect moving vehicles successfully and to remove noise signatures to obtain correct vehicle signatures. However, they argued that future research work and sufficient funds on this technology will be necessary to deploy this new train detection technology at HRGCs as an alternate train warning system.

Video Detection Technology

Video image detection techniques have been used at HRGC to detect train movements (Appiah and Rilett 2008; Chen 2015; Forsberg 2012; Martin et al. 2004; Tian 2003). Video image detection technologies involve an image processor, which extracts necessary information from a video footage. These detection devices have capabilities to obtain speed, occupancy, count, and the presence or absence of vehicular objects. To have information of vehicles or trains at a specified area, the system sets one or several detection boundaries within the image range of the camera. Using real-time image processing algorithms, the system detects any changes within the boundaries for desired objectives. A common operational issue of this technology is occlusion, which results in missed detection, false detection, and increased detector presence time (Tian 2003). Occlusion takes place because of the parallax effect in video detection systems, as shown in Figure 2.4. As two different objects are overlapped at one detection zone, the video detection system may not correctly identify the two as separate objects due to the overlapping (e.g., one behind the other cannot be recognized by the detector). To remove false detection, the position of detection cameras should be high enough to secure both rail tracks in a detection zone (Tian 2003).

A study investigated various environmental factors that may affect the performance of the accuracy of the video detection system in eight locations in Utah (Martin et al. 2004). Data were obtained under different weather (clear, snow, rain, and fog), and light conditions (day, night, and dusk). The study found that the detection accuracy was the highest during day and dusk conditions, having approximately 87 percent of the correct detection rate. The rate, however, was reduced during the nighttime (73.4 percent) and under severe weather conditions (81.3 percent). The study recommended that the detailed installation information of video detection equipment be carefully reviewed before deploying. Such information includes proper placement of cameras, enough background lighting acquisition, camera focus settings, and field of view calibration to minimize false detection.

What is a Fusion Sensor?

A Fusion Sensor, also referred to a sensor fusion, uses multiple sensors along with logic to deduce sensed result based on the strengths and weaknesses of individual sensors. Several detection technologies can be combined to provide additional information on trains as well as highway vehicle status. Multiple detectors combined in a system have significant potential of generating more detailed and accurate train information that may be a source of future traffic control systems (Reiff et al. 2001, 2003). The FRA, Transportation Technology Center, Inc. (TTCI), and the John A. Volpe National Transportation Systems Center (Volpe Center) conducted a study to evaluate five combined systems identified as System 1, 2, 3, 4, and 6 (Reiff et al. 2003). Detailed technical information of each system, as provided by the system vendors, appears in the original report (ibid.). The authors concluded that the evaluated prototype systems did not always provide satisfactory train and highway vehicle detection and recommended future studies on these technologies for improved performance (ibid).

LinqThingz has made great improvements on concepts explored in the past and have leveraged significant advancement in sensor, cloud and machine learning. The advent of commercial automated vehicles has facilitated the engineering improvement of traffic sensors and the volume production have significantly dropped the cost from two decades ago. Linqthingz leverages RADAR, LIDAR, Camera, Magnetometer, Audio, RF, Infrared, Environmental Sensors and third-party sensor data into a single package.

LinqThingz uses sensor fusion along with a trained neural network (AI) to differentiate train, cross gate, vehicles, people and other objects in order to maximize Receiver Operation Curve (ROC). In other words, maximize true positives, maximize true negatives, minimize false positives, minimize false negatives. The is an example of bio-mimicry. Animals typically use the combination of multiple sensors (sight, sound, smell, touch, etc.) to deduce a result.

LinqThingz also provides a distributed AI system that takes data from multiple sensors to understand motion of trains and other object resulting in Predictive Mobility. LinqThingz Predictive Mobility Platform provide information to control systems, to cloud based transportation management systems through our API and to human users through web and mobile applications.