Site-based calibration of object detection parameters
Systems and methods for site-based calibration of object detection values, such as for surveillance video cameras, are described. Video data from a video image sensor may be processed using an object detector to determine object data and a confidence score for a detected object. The object data and confidence score may be post-processed to apply calibration values based on the camera location to one or more parameters used for determining detection events. Event notifications may be sent for detection events. The calibration values may be determined from a calibration period where a verification object detector is used to verify the object detections and failure analysis is applied to determine calibration values for the camera location.
1. A system, comprising:
a video image sensor; and
a surveillance controller configured to:
receive video data from the video image sensor, wherein the video data includes a video stream of video frames captured by the video image sensor at a camera location;
determine, during a calibration period, at least one calibration value for the camera location by:
determining, using a field object detector based on a first object detection model, field object data for a plurality of calibration objects;
determining, based on an object verification detector using a second object detection model, verification object data for the plurality of calibration objects, wherein:
the first object detection model has a first accuracy value for objects of interest;
the second object detection model has a second accuracy value for objects of interest; and
the second accuracy value is higher than the first accuracy value;
comparing the field object data from the field object detector to the verification object data for the plurality of calibration objects;
determining, based on the comparison of the field object data to the verification object data, whether the verification object data matches the field object data;
determining, responsive to the verification object data including objects of interest not in the field object data, missed object detection conditions;
determining, responsive to the field object data including objects of interest not in the verification object data, false object detection conditions; and
determining, based on the comparison of the field object data and the verification object data, the at least one calibration value to compensate for a failure condition of the first object detection model, wherein the failure condition is selected from the missed object detection conditions and the false object detection conditions;
determine, using the field object detector during an operating period:
object data for a detected object in the video data; and
a confidence score for the detected object;
post-process, during the operating period, the object data and the confidence score to determine a calibrated detection event, wherein post-processing includes applying the at least one calibration value for the camera location to determine the calibrated detection event; and
send, responsive to the calibrated detection event, an event notification to a video surveillance application.
2. The system of claim 1 , further comprising:
a calibration controller comprising the object verification detector and configured to:
receive, during the calibration period, calibration video data from the video image sensor at the camera location;
receive, during the calibration period, the field object data for the plurality of calibration objects;
determine, using the object verification detector, the verification object data for the plurality of calibration objects in the calibration video data;
determine, based on the comparison of the field object data and the verification object data:
the false object detection conditions determined by the field object detector; and
the missed object detection conditions determined by the field object detector; and
determine, based on the false object detection conditions and the missed object detection conditions, the at least one calibration value for post-processing the object data and the confidence score for the detected object during the operating period.
3. The system of claim 2 , further comprising:
a video camera housing, wherein the video camera housing encloses:
the video image sensor;
the surveillance controller; and
a network interface configured to communicate with a network, wherein the surveillance controller is further configured to:
send, to the calibration controller and over the network:
the field object data; and
the calibration video data from the video image sensor; and
receive, from the calibration controller and over the network, the at least one calibration value determined based on the calibration period.
4. The system of claim 2 , wherein:
the second object detection model of the object verification detector includes:
a motion model configured to determine a bounding box of moving objects in the calibration video data; and
an image classifier configured to detect an object of interest in the calibration video data; and
the calibration controller is further configured to:
compare bounding boxes of moving objects to detected objects from the field object data;
determine whether the object of interest is present in unmatched bounding boxes; and
determine, responsive to the image classifier detecting the object of interest in unmatched bounding boxes, the missed object detection conditions.
5. The system of claim 2 , wherein:
the second object detection model of the object verification detector includes an image classifier configured to detect an object of interest in the calibration video data; and
the calibration controller is further configured to:
determine whether the object of interest is absent in the field object data; and
determine, responsive to the image classifier not detecting the object of interest in the field object data, the false object detection conditions.
6. The system of claim 2 , wherein:
the second object detection model of the object verification detector includes:
a tracking model configured to determine a bounding box of an object of interest in successive frames in the calibration video data; and
an image classifier configured to detect the object of interest in the calibration video data; and
the calibration controller is further configured to:
compare bounding boxes in successive frames to detected objects from the field object data;
determine whether the object of interest is present in unmatched bounding boxes; and
determine, responsive to the image classifier detecting the object of interest in unmatched bounding boxes, the missed object detection conditions.
7. The system of claim 2 , wherein:
the second object detection model is configured to determine
verification confidence scores for the detected objects; and
the calibration controller is further configured to execute, for the surveillance controller:
comparing the field object data to the verification object data for the detected objects;
determining, based on the comparison of the field object data to the verification object data, whether the verification object data matches the field object data;
determining, responsive to the verification object data including objects of interest not in the field object data, the missed object detection conditions; and
determining, responsive to the field object data including objects of interest not in the verification object data, the false object detection conditions.
8. The system of claim 2 , wherein:
the calibration controller is further configured to selectively receive, during the operating period:
the video data from the video image sensor; and
the field object data;
the surveillance controller is further configured to evaluate verification conditions for sending data to the calibration controller; and
the verification conditions are selected from:
object verification scores from the field object detector meeting a verification threshold; and
periodic sampling of the video stream.
9. The system of claim 2 , wherein:
the calibration controller is further configured to:
aggregate the missed object detection conditions;
aggregate the false object detection conditions;
determine a set of adjustment conditions associated with the camera location; and
determine, based on failure analysis, a plurality of calibration values corresponding to a plurality of adjustment conditions in the set of adjustment conditions; and
the surveillance controller is further configured to:
determine a current adjustment condition at the camera location; and
determine, using the current adjustment condition at the camera location, the at least one calibration value from the plurality of calibration values.
10. The system of claim 1 , wherein the surveillance controller is further configured to:
apply the at least one calibration value to a confidence score threshold; and
compare the confidence score for the detected object to the confidence score threshold to determine the calibrated detection event.
11. The system of claim 1 , wherein the surveillance controller is further configured to:
apply the at least one calibration value to a bounding box size threshold; and
compare a bounding box size for the detected object to the bounding box size threshold to determine the calibrated detection event.
12. The system of claim 1 , wherein the surveillance controller is further configured to:
apply the at least one calibration value to an intersection-over-union threshold; and
compare an intersection-over-union value for the detected object to the intersection-over-union threshold to determine the calibrated detection event.
13. The system of claim 1 , wherein the surveillance controller is further configured to:
apply the at least one calibration value to a motion model sensitivity threshold; and
compare a motion value for the detected object to the motion model sensitivity threshold to determine the calibrated detection event.
14. The system of claim 1 , wherein:
the surveillance controller is further configured to:
determine a set of adjustment conditions for the camera location;
determine a plurality of calibration values for determining the calibrated detection event;
map the set of adjustment conditions to the plurality of calibration values;
determine a current adjustment condition at the camera location; and
determine, using the current adjustment condition to index the set of adjustment conditions, the at least one calibration value from the plurality of calibration values; and
the set of adjustment conditions are selected from condition types including:
relative object size;
location in video frame;
time;
light condition; and
weather condition.
15. A computer-implemented method, comprising:
receiving video data from a video image sensor, wherein the video data includes a video stream of video frames captured by the video image sensor at a camera location;
determining, during a calibration period, at least one calibration value for the camera location by:
determining, using a field object detector based on a first object detection model, field object data for a plurality of calibration objects in calibration video data from the video image sensor;
determining, based on an object verification detector using a second object detection model, verification object data for the plurality of calibration objects in the calibration video data, wherein:
the first object detection model has a first accuracy value for objects of interest;
the second object detection model has a second accuracy value for objects of interest; and
the second accuracy value is higher than the first accuracy value;
comparing the field object data from the field object detector to the verification object data for the plurality of calibration objects;
determining, based on the comparison of the field object data to the verification object data, whether the verification object data matches the field object data;
determining, responsive to the verification object data including objects of interest not in the field object data, missed object detection conditions;
determining, responsive to the field object data including objects of interest not in the verification object data, false object detection conditions; and
determining, based on the comparison of the field object data and the verification object data, the at least one calibration value to compensate for a failure condition of the first object detection model, wherein the failure condition is selected from the missed object detection conditions and the false object detection conditions;
determining, using the field object detector during an operating period:
object data for a detected object in the video data; and
a confidence score for the detected object;
post-processing, during the operating period, the object data and the confidence score to determine a calibrated detection event, wherein post-processing includes applying the at least one calibration value for the camera location to determining the calibrated detection event; and
sending, responsive to the calibrated detection event, an event notification to a video surveillance application.
16. The computer-implemented method of claim 15 , further comprising:
determining, using the second object detection model, a bounding box of moving objects in the calibration video data;
detecting, using an object classifier, an object of interest in the calibration video data; and
comparing bounding boxes of moving objects to detected objects from the field object data;
determining whether the object of interest is present in unmatched bounding boxes; and
determining, responsive to detecting the object of interest in unmatched bounding boxes, the missed object detection conditions.
17. The computer-implemented method of claim 15 , further comprising:
detecting, by the second object detection model, an object of interest in the calibration video data;
determining whether the object of interest is absent in the field object data; and
determining, responsive to not detecting the object of interest in the field object data, the false object detection conditions.
18. The computer-implemented method of claim 15 , further comprising:
aggregating, during the calibration period, the missed object detection conditions;
aggregating, during the calibration period, the false object detection conditions;
determining a set of adjustment conditions associated with the camera location;
determining, based on failure analysis of aggregate missed object detection conditions and aggregate false object detection conditions, a plurality of calibration values corresponding to a plurality of adjustment conditions in the set of adjustment conditions;
determining, during the operation period, a current adjustment condition at the camera location; and
determining, using the current adjustment condition at the camera location, the at least one calibration value from the plurality of calibration values.
19. The computer-implemented method of claim 15 , further comprising:
applying the at least one calibration value to a post-processing threshold; and
comparing a corresponding value for the detected object to the post-processing threshold to determine the calibrated detection event, wherein the post-processing threshold is selected from:
a confidence score threshold;
a bounding box size threshold;
an intersection-over-union threshold; and
a motion model sensitivity threshold.
20. A video camera, comprising:
a video image sensor;
means for receiving video data from the video image sensor, wherein the video data includes a video stream of video frames captured by the video image sensor at a camera location;
means for determining, during a calibration period, at least one calibration value for the camera location by:
determining, using a field object detector based on a first object detection model, field object data for a plurality of calibration objects;
determining, based on an object verification detector using a second object detection model, verification object data for the plurality of calibration objects, wherein:
the first object detection model has a first accuracy value for objects of interest;
the second object detection model has a second accuracy value for objects of interest; and
the second accuracy value is higher than the first accuracy value;
comparing the field object data from the field object detector to the verification object data for the plurality of calibration objects;
determining, based on the comparison of the field object data to the verification object data, whether the verification object data matches the field object data;
determining, responsive to the verification object data including objects of interest not in the field object data, missed object detection conditions;
determining, responsive to the field object data including objects of interest not in the verification object data, false object detection conditions; and
determining, based on the comparison of the field object data and the verification object data, the at least one calibration value to compensate for a failure condition of the first object detection model, wherein the failure condition is selected from the missed object detection conditions and the false object detection conditions;
means for determining, using the field object detector during an operating period:
object data for a detected object in the video data; and
a confidence score for the detected object;
means for post-processing, during the operating period, the object data and the confidence score to determine a calibrated detection event, wherein post-processing includes applying the at least one calibration value for the camera location to determining the calibrated detection event; and
means for sending, responsive to the calibrated detection event, an event notification to a video surveillance application.