| CPC G06Q 50/30 (2013.01) [G06F 18/214 (2023.01); G06F 18/23 (2023.01); G06V 20/10 (2022.01); G06V 20/582 (2022.01); G06V 20/584 (2022.01); G06V 2201/07 (2022.01); G06V 2201/08 (2022.01); G06V 2201/10 (2022.01)] | 3 Claims |
|
1. A processor implemented method for detecting on-street parking violations, comprising:
receiving, via one or more first hardware processors associated with an edge based electronic device, a video stream of a scene during a trip of a vehicle, the video stream captured by using a media capturing device embodied in the edge based electronic device mounted in the vehicle, and calibrating the media capturing device prior to capturing the video stream of the scene,
enabling, via the one or more first hardware processors, a lightweight information technology application to capture videos of one or more streets by the media capturing device and identifying one or more feature sets, the one or more feature sets being used as inputs for one or more high level inferencing tasks;
processing, via the one or more first hardware processors, the video stream in real time to identify a plurality of objects of interest and a meta-information associated with the plurality of objects of interest, the plurality of objects of interest comprises one or more no-parking signage and one or more vehicles in a predetermined area around the one or more no-parking signage, the meta-information associated with the plurality of objects of interest stored in a repository associated with the edge based electronic device in form of a historian, wherein the meta-information includes time of detection of the plurality of the objects of interest, location of the media capturing device at the time of detection, and distance of the media capturing device from the plurality of the objects of interest, and wherein processing the video stream to identify the plurality of objects of interest comprises locating, using convolutional neural network (CNN) models, object detection instances in the video stream, and identifying the object detection instance as one of a parked vehicles and a no-parking signage and on detection of a no-parking signage:
triggering estimation of a separation distance between the signage from the one or more parked vehicles; and
flagging a parking violation if the distance of the one or more parked vehicles is determined to be less than or equal to a second predetermined threshold distance span on either side of the signage;
performing, via the one or more first hardware processors, object detection by using an object detector model configured by a single-shot multi box detector (SSD) on a MobileNet V2 deep architecture, wherein performing object detection comprises training image data with one or more bounding box hypotheses and selecting a category for a fixed set of bounding boxes by providing a highest prediction score;
identifying, via the one or more first hardware processors, an exhaustive feature set, the exhaustive feature set being subsequently used as inputs to the CNN models in a fully connected neural network for high level inferencing tasks;
performing, via the one or more first hardware processors, transfer learning, wherein a pre-trained model is reused as a starting point for training a new model for one or more tasks, and wherein the pre-trained model is re-used by combining with a custom curated dataset, the custom curated dataset being curated by including one or more detected parking violations;
computing, via the one or more first hardware processors, absolute locations of the plurality of objects of interest, wherein obtaining the absolute locations of the plurality of objects of interest comprises:
determining, for two or more consecutive object detection instances in the video stream, whether locations values of the media capturing device along x-coordinate and y-coordinate are changing, and
computing, on determination of changing of locations values of the media capturing device along at least one of x-coordinate and y-coordinate, the absolute locations based on relative location values of the media capturing device from the detected objects, wherein computing the absolute locations based on the determination of relative locations values of the media capturing device with respect to one or more vehicles, comprises:
appending distance of the media capturing device to the value of x-coordinate to obtain the absolute location when the value of x-coordinate is determined to change;
appending distance of the media capturing device to the value of y-coordinate to obtain the absolute location when the value of y-coordinate is determined to change; and
appending (distance/2) of the media capturing device to the value of the x-coordinate and the y-coordinate to obtain the absolute location when the value of x-coordinate and the y-coordinate are determined to change;
updating, via the one or more first hardware processors, the historian with the absolute locations of the one or more no-parking signage and the one or more vehicles to obtain an updated historian;
filtering the updated historian, via the one or more first hardware processors, to obtain a filtered historian, the filtering is performed to (1) determine a set of unique objects of interest from amongst the plurality of objects of interest, and (2) identify each vehicle from amongst the one or more vehicles as one of parked vehicle and moving vehicle , wherein the set of unique objects of interest and meta-information associated with the parked vehicle is stored in the filtered historian, wherein the set of unique objects are identified from amongst the plurality of objects of interest based on a sequential clustering of the absolute locations of the plurality of objects of interest, wherein the sequential clustering of the absolute locations of the plurality of objects of interest comprises:
computing a distance between consecutive object locations associated with the objects of interest; and
clustering the consecutive object locations as belonging to an object of interest from amongst the plurality of objects of interest when the consecutive object locations are spaced at a distance less than a first predetermined threshold distance, wherein the first predetermined threshold distance is obtained from a parking violation data that is aggregated in at least one of the edge based electronic device and a cloud server; and
wherein identifying each vehicle from amongst the one or more vehicles as one of parked vehicle and moving vehicle comprises differentiating between parked vehicles and moving vehicles from amongst the plurality of objects of interest, wherein differentiating comprises:
determining one of, during the trip of the vehicle:
whether the distance between the location of the media capturing device and the vehicle is decreasing during the two or more consecutive object detection instances, and
whether the bounding box dimension around the identified vehicle in the video stream is increasing during the two or more consecutive object detection instances; and
tagging the vehicle as the parked vehicle when:
the distance between the media capturing device location and the vehicle decreases during the two or more consecutive object detection instances, or
the bounding box dimension of the identified vehicle increases during the two or more consecutive object detection instances;
applying meta-tagging, via the one or more first hardware processors to the plurality of parked vehicles identified on the video stream with locations and time of detection of the plurality of parked vehicles;
triggering estimation, via the one or more first hardware processors, of a separation distance between the signage from one or more parked vehicles, wherein if the distance of the one or more parked vehicles is determined to be less than or equal to a second predetermined threshold distance span on either side of the signage, then a parking violation is flagged;
and
determining, via the one or more first hardware processors, the parking violation at scale based on a comparison of the filtered historian with the parking violation data associated with a prescribed parking violation policy, wherein determining the parking violations comprises sending the meta-information associated with parked vehicles to a cloud server, the cloud server comprising one or more second hardware processors configured to determine the parking violations based on the meta-information, wherein the parking violations comprises detecting parking violation conditions including at least one of occluded, blurred, distorted, and broken sign-boards.
|