The features of the sensors we developed include the capability for areal monitoring of the measurement area, and the capability for monitoring a 10-m X 100-m area (roughly three lanes). As a result, they are distinguished by the ability to determine what locations within the area appear to be most problematic.

  These sensors also display the five conditions of dry, wet, water film, snow cover, and freezing. We have verified on actual road surfaces that they can determine these conditions with a certain degree of apparent certainty. In order to determine freezing, however, they require a road surface thermometer in addition to image information. (Figure 4)

Figure 4

  First of all, we have verified 90% accuracy for the five conditions. The degree of accuracy of the sensors on actual road conditions is expressed by the sensing accuracy rate. The extent to which the sensor results of dry, wet, and so on match actual conditions is expressed by the sensor guess right ratio. Both achieve levels close to the provisional target value of 90%. Unfortunately, the Miyako test site did not experience any freezing, so the freezing data on this graph was taken from the Nakayama Pass in Hokkaido, which uses the same system. Given that the system can determine these five conditions, the question then becomes how it would be used for road management. (Figure 5)

Figure 5

  It may appear that these five conditions are simply divided between wet and water film. However, I think you will find two general points of entry.

   First of all, water film (when water forms a layer on the road surface) has a greater heat content than the wet condition. Therefore it will be later to freeze when subjected to the same change in air temperature. I expect that this characteristic could be used as an index for mobilization readiness and road surface management.

  Moreover, raising the number of conditions from four to five is also something of an improvement in terms of system accuracy. This seems likely to provide an avenue for improving the accuracy of freezing forecasting.

  Here are images of areal measurement, with the raw image on the left and a processed image on the right. The area is divided into a mesh that is about 50 cm square at the smallest points. With the conventional method on the left, only a spot area such as this is visible. Moreover, because it uses four conditions, the display will show a wet condition even when there is water film. (Figure 6)

Figure 6

  This may be all right when the road surface is homogeneous, as with water film. As that surface gradually dries, however, the water film develops a growing number of wet areas within it, and then dry and wet areas. The conventional sensors will identify this mixed condition as dry based on its spot area. We sometimes find a scenario in which the snow cover condition is a mixture of wet, water film, and snow cover conditions due to wheel tracks. Conventional sensors will determine this as a snow cover condition if the spot they happen to be monitoring has accumulated snow. This determination is on the relatively safe side. Given that condition, however, if the temperature drops slightly so that the wet and water film portions start changing to freezing, as shown in Figure 6, then the conventional sensor may fail to detect the most fearsome condition.

  I would like to leave it up to the road administrators to judge how effective the system may be. However, I want to point out that the ability to distinguish this worst condition within such an area is one feature of the system.

  Actual road surfaces have a variety of conditions coming together across the expanse of an area. The system is distinguished by the capability for picking up such places visually. I would be very pleased if you could actually enjoy putting this to use in various ways.

  Here we see five cameras in use. This should be advantageous in terms of installation space and cost, and we have studied the use of a single unit to process input from multiple cameras. The horizontal axis shows the number of cameras connected, and the vertical axis shows the relative sensing accuracy rate. In other words, a single camera gives the value 1, and then the graph goes on to show how accuracy diminishes as two, three, four, and five cameras are connected. The input can be processed fairly accurately with up to five cameras. (Figure 7)

Figure 7

  Here the internal process is outputting data at one-minute intervals. However, the data for the ten-minute period is observed constantly, and a running average is kept. Where 11 items of data were processed for a single camera, this becomes six items for two cameras, four items for three cameras. The data count goes down, and the accuracy diminishes as the number of cameras increases. We found, therefore, that the system is capable of suitable detection accuracy with up to about five cameras.

  Comparison of the data with one camera and data with five cameras shows that they can assure roughly equal performance. (Figure 8)

Figure 8

  Next, you will see that we have verified how far the system can stand up to the use of swivel and zoom cameras, taking into account the familiar similarity to already installed cameras for road administrators (Figure 9).

Figure 9

  The left side shows a fixed camera. The right shows a camera that has preset conditions after it has performed swivel and zoom and then been returned to the initial setting. The performance has largely held steady close to the expected level.

  The system can make good use of the cameras already installed, so that there is no need for further facilities or placement of cameras on-site. The system can also process imagery from cameras with viewing angles close to those now in use. More than that, we developed a system that can process data from five cameras, including the use of swivel and zoom. In terms of the technology, we find that the outlook is very clear.

  In developing technology for detecting individual vehicle behavior in other words, detecting evasive maneuvering, standing, and slow vehicles we consider that we have also developed technology for early discovery of road events such as falling rock, collapsing slope faces, and so on. (Figures 11, 12, and 13)

Figure 11

Figure 12

  Discovery of events such as falling rock, for example, would of course be best accomplished by sensors that could pick up rocks that fall on the road surface. Unfortunately, however, present technology does not easily allow us to pick up falling rocks about 10 cm in size 100 m ahead using images from visible-light cameras. Therefore, we think that events such as falling rock could be picked up instead, though somewhat indirectly, by making out the evasive maneuvering of vehicles caused by falling rock on the road surface and so on.

  On a slightly larger scale, when we are dealing with such large-scale events as collapsing slope faces, the road will become impassable to vehicles, and it will probably be possible to pick up standing, slow, and congested vehicles, and so on. An event such as a collapsed slope face that fills the entire screen will result in traffic volume being reduced from the usual one or two vehicles every five minutes to none at all in ten minutes or thirty minutes, and we think it possible that the event could be discovered early in this way.

  The flow of detection would follow from maneuvers that vehicles take to avoid fallen rocks when they appear on the roadway. The parameters of so many vehicles in so many minutes could be set in consultation with managers, taking into account traffic volume at that location, the road shape, and so on. Sensors would then pick up evasive maneuvering in accordance with the definition of evasive maneuvering for each site. The sensors would issue a report based on the detection results, and administrators would then be able to proceed with countermeasures, by visually confirming the reported information, or going to that location if necessary. (Figure 13)

Figure 13

  Finally, there is the detecting technology used in road sensors. Since this kind of detecting technology and identification of behavior have become possible, it has also become technically possible to detect events such as evasive maneuvering, standing vehicles, and slow vehicles while at the same time recording the images before and after the events. We expect that, as a result, it should be possible to record events in their entirety, including accidents. The ability to record these images themselves, and to accumulate individual vehicle data such as speed and position together with the recorded data, may open up possibilities for wide-ranging use in the management area, such as, depending on the case, reevaluation of design standards, and so on. (Figure 14)

Figure 14

  Today I have introduced visible image type road condition and road surface sensor that have been brought more or less to the intended technical level. There are also millimeter-wave radar type sensor and laser type sensor, in addition to these, and I hope that we will be able to propose various uses for those sensors when the time comes.