Object recognition device and vehicle control system

A vehicle control system comprising:

• an object recognition device configured to execute object recognition processing that recognizes an object around a vehicle;
• and
• a vehicle travel control device configured to execute vehicle travel control that controls travel of the vehicle based on a result of the object recognition processing, wherein
• the vehicle includes a LIDAR (Laser Imaging Detection and Ranging),
• a LIDAR point cloud is a set of detected points detected by the LIDAR,
• the object recognition device includes:
• a memory device in which LIDAR measurement information indicating a detected position of each of the detected points, position attitude information indicating an absolute position and an attitude of the vehicle, and terrain map information indicating an absolute position of a road surface are stored;
• and
• a processor configured to execute the object recognition processing,
• the object includes a fallen object on the road surface and a tracking target other than the fallen object, and
• in the object recognition processing, the processor:
• sets a space having a first height from the absolute position of the road surface as a noise space, based on the terrain map information;
• classifies the LIDAR point cloud into a first point cloud included in the noise space and a second point cloud not included in the noise space, based on the LIDAR measurement information and the position attitude information;
• executes first extraction processing that extracts a fallen object candidate being a candidate for the fallen object, based on the first point cloud;
• executes second extraction processing that extracts a tracking target candidate being a candidate for the tracking target, based on the second point cloud;
• determines whether or not horizontal positions of the fallen object candidate and the tracking target candidate are consistent with each other;
• integrates the fallen object candidate and the tracking target candidate whose horizontal positions are consistent with each other, to be the tracking target;
• and
• recognizes the fallen object candidate not integrated into the tracking target as the fallen object, wherein:
• a road surface point cloud is the LIDAR point cloud representing the road surface,
• the first point cloud includes the road surface point cloud and a fallen object point cloud other than the road surface point cloud,
• in the first extraction processing, the processor:
• executes separation processing that separates the first point cloud into the road surface point cloud and the fallen object point cloud, based on the LIDAR measurement information or the terrain map information;
• and
• extracts the fallen object candidate based on the fallen object point cloud
• regarding the first point cloud belonging to a same layer, there is a difference between the detected position of the fallen object point cloud and the detected position of the road surface point cloud,
• a feature quantity is a parameter characterizing the difference,
• in the separation processing, the processor:
• calculates the feature quantity of each of the detected points of the first point cloud, based on the LIDAR measurement information on the first point cloud;
• and
• separates the first point cloud into the road surface point cloud and the fallen object point cloud, based on the feature quantity
• a fallen object likelihood is likelihood that the fallen object candidate is the fallen object,
• the processor further executes likelihood calculation processing that calculates the fallen object likelihood after extracting the fallen object candidate,
• a separation criterion for separating the road surface point cloud and the fallen object point cloud is predetermined,
• in the likelihood calculation processing, the processor calculates the fallen object likelihood based on the separation criterion and the feature quantity regarding the fallen object point cloud corresponding to the fallen object candidate, and
• the fallen object likelihood becomes higher as the feature quantity regarding the fallen object point cloud corresponding to the fallen object candidate is further away from the separation criterion.

An object recognition device that recognizes an object around a vehicle,

• the vehicle comprising a LIDAR (Laser Imaging Detection and Ranging),
• a LIDAR point cloud being a set of detected points detected by the LIDAR,
• the object recognition device comprising:
• a memory device in which LIDAR measurement information indicating a detected position of each of the detected points, position attitude information indicating an absolute position and an attitude of the vehicle, and terrain map information indicating an absolute position of a road surface are stored;
• and
• a processor configured to execute object recognition processing that recognizes the object around the vehicle, wherein
• the object includes a fallen object on the road surface and a tracking target other than the fallen object, and
• in the object recognition processing, the processor:
• sets a space having a first height from the absolute position of the road surface as a noise space, based on the terrain map information;
• classifies the LIDAR point cloud into a first point cloud included in the noise space and a second point cloud not included in the noise space, based on the LIDAR measurement information and the position attitude information;
• executes first extraction processing that extracts a fallen object candidate being a candidate for the fallen object, based on the first point cloud;
• executes second extraction processing that extracts a tracking target candidate being a candidate for the tracking target, based on the second point cloud;
• determines whether or not horizontal positions of the fallen object candidate and the tracking target candidate are consistent with each other;
• integrates the fallen object candidate and the tracking target candidate whose horizontal positions are consistent with each other, to be the tracking target;
• and
• recognizes the fallen object candidate not integrated into the tracking target as the fallen object, wherein:
• a road surface point cloud is the LIDAR point cloud representing the road surface,
• the first point cloud includes the road surface point cloud and a fallen object point cloud other than the road surface point cloud,
• in the first extraction processing, the processor:
• executes separation processing that separates the first point cloud into the road surface point cloud and the fallen object point cloud, based on the LIDAR measurement information or the terrain map information;
• and
• extracts the fallen object candidate based on the fallen object point cloud
• regarding the first point cloud belonging to a same layer, there is a difference between the detected position of the fallen object point cloud and the detected position of the road surface point cloud,
• a feature quantity is a parameter characterizing the difference,
• in the separation processing, the processor:
• calculates the feature quantity of each of the detected points of the first point cloud, based on the LIDAR measurement information on the first point cloud;
• and
• separates the first point cloud into the road surface point cloud and the fallen object point cloud, based on the feature quantity
• a fallen object likelihood is likelihood that the fallen object candidate is the fallen object,
• the processor further executes likelihood calculation processing that calculates the fallen object likelihood after extracting the fallen object candidate,
• a continuity represents a degree to which a position of the fallen object candidate is continuous over time, and
• in the likelihood calculation processing, the processor:
• calculates the continuity based on an extraction history of the fallen object candidate;
• increases the fallen object likelihood, when the continuity is equal to or higher than a continuity threshold;
• and
• decreases the fallen object likelihood, when the continuity is lower than the continuity threshold.