340.

FIG. 4 illustrates a process 400 for in-vehicle detection of potential false detections, according to an embodiment of the present disclosure. Process 400 may not rely on external or crowdsourced traffic sign data, but rather may utilize internal sensors, for example a front camera module. Thus, at step 410 a sign may be detected utilizing the front camera module or other types of image capture devices including but not limited to light detection and ranging (Lidar), radar, or the like. While this example illustrates a speed limit traffic sign, the type of traffic sign is not limited to speed but may be of other types of traffic signs. At step 420 a determination may be made as to whether the host vehicle's speed is within a threshold amount of the detected speed captured by the front facing sensor. If the host vehicle speed is close and within the threshold amount of the detected speed limit, then the situation may be classified as a low risk at step 430. However, if the host vehicle speed is not close to the detected speed limit, then at step 440 a determination may be made as to whether the host vehicle speed is less than the detected speed limit. If it is less then a determination may be made at step 450 as to whether the host vehicle is part of a road congestion with slow moving traffic and if so, then the associated risk flag may be determined to be a low risk at step 430. However, if the host vehicle speed is less than the detected speed limit at step 440 and there is no slow-moving traffic detected at step 450, then the situation may be determined to be a high risk at step 460. And similarly, if the host vehicle speed is higher than the detected speed limit then the situation may also be flagged as a high risk at step 460.

FIG. 5 illustrates the processing of high-speed vehicle telemetry 500 utilizing crowdsourced data, according to an embodiment of the present disclosure. At step 510, crowdsourced traffic sign data may be collected. Such data may include multiple types of traffic signs and may also include the traffic sign's associated location. Crowdsourced traffic sign data may be collected over time and pertain to one or more roadways. At step 520, based on the collected crowdsourced data in step 510, a determination may be made as to the reported traffic sign in question. In step 525 there may also be an associated confidence score associated with the reported traffic sign. There may also be, at step 530, a mixture of correct detection and one or more wrong detections associated with the reported traffic sign. The output of both step 525 and step 530 may be presented, in step 540, as the inferred traffic sign based on the processed high speed vehicle telemetry from step 535. The processed high speed vehicle telemetry may consist of multiple filters and distribution methods of raw high speed vehicle telemetry data. Such filtering may include time and spatial filtering to select free-flow traffic data where the average vehicle speed may be measured during low traffic volume periods and/or also be filtering out high congestion regions. In addition, a traffic sign filter may eliminate certain data by category type such as deleting low values of speed limits, keeping values associated with a stop sign, and also keeping values associated with a yield sign. Processing the high speed vehicle telemetry raw data may also include the use of distribution methods such as a histogram, density plot, and kernel density estimation to find the values range where most of the data exists, obtaining the minimum and maximum values of the highest bin, and determining an average of the highest bin. Examples of distribution methods will be further discussed in FIG. 7.

Thus, at step 540 the result of the high-speed vehicle telemetry may result in an improved traffic sign confidence score in step 545 and an ability to identify and/or rectify wrong traffic sign detections in step 550.

FIG. 6 is a further discussion of processing 600 of crowdsourced traffic sign data, according to an embodiment of the present disclosure. The processing 600 of high speed vehicle telemetry may begin at step 610 with obtaining traffic sign type and location data. Further, the traffic sign type and location data may then be grouped by the type of traffic sign, for example, by speed, stop, yield, etc. In addition, the data may also be grouped by an edge identification or road identification or designation. At step 620 the data may then be curated by filtering out low data based on a type of traffic sign, for example, the removal of zero speed data from a speed limit type of traffic sign.

At step 630 a time and spatial filtering may be used to select free-flow traffic data. For example, an average vehicle speed may be measured during low-volume periods of time, thus filtering our high congestion periods. Further, the high congestion periods may be defined by times when the traffic in one or more directions and segments or portions of a roadway is typically operating below free-flow speed, for example between 06:00-10:00 and between 15:00 and 18:00. These times are simply arbitrary and are not meant to be restrictive.

Next, a spatial filtering may be used to select a most appropriate set of telemetry data sample for analysis. For example, when processing highway telemetry data, samples near an exit/merge ramp may be filtered out. Similarly, for a city/residential road, samples near an intersection may be filtered out. This may be performed by using OSM (or other) road topology information such as vertices to identify where roads intersect and filter out samples using a distance threshold. Another method may include the ability to select vehicle telemetry samples within a distance threshold from the detected traffic sign. In addition, behavior near a posted traffic sign may be observed to further quantify the effect of the traffic sign and further confirm a legend of the associated traffic sign.

At step 640 distribution methods may be applied to the data and may include the use of histograms, density plots, kernel density estimation and the like. Such methods may include finding value ranges that contain the majority of data, for example the mode of a histogram. Then, a determination of the minimum and maximum values of the highest bin may be made, an average of the highest bin obtained, and a confidence score estimated.

In step 650, based on the results of the distribution methods in step 640, a high-speed vehicle telemetry inferred traffic sign legend, for example, speed, stop, or yield, may be determined. The inferred traffic sign data of step 650 may then be further processed in step 660, step 670, and step 680. Step 660 flag condition may identify and flag unexpected traffic signs and objects. Step 670 may also determine false detections using vehicle sensors and further aggregate the incorrect detections. Step 680 may also perform a de-duplication process based on the inferred traffic sign data, which will be further discussed in FIG. 8.

Then, at step 690, based on the determinations of step 660, step 670, and step 680, a final traffic sign value or legend may be identified as the most likely traffic sign legend.

FIG. 7 is an estimation 700 of a confidence score of high-speed vehicle telemetry data based on data distribution, according to an embodiment of the present disclosure.