Verkada Inc.

Yi Xu

Yunchao Gong

Mayank Gupta

Naresh Nagabushan

Yuanyuan Chen

Zixiao (Shawn) Wang

Qiang (Kevin) Fu

Xia Yang

Rishabh Goyal

Suraj Arun Vathsa

Song Cao

Hao Nan

Thantham Madan

Yingjie Shen

Anders Brekke Holden

Bhavna Sud

Nitish Poddar

Michele Casertano

David Mwaura

Filip Kaliszan

Shiqi Wang

A trajectory-based object search process includes receiving a user request specifying a geographic region of interest and/or a time period of interest. A set of video-derived object trajectories is identified based on the user request. At least one event is identified based on the set of video-derived object trajectories, by processing each video-derived object trajectory individually or by aggregating video-derived object trajectories, from the set of video-derived object trajectories, that overlap with respect to at least one feature. For each event, a preview image is selected that includes a thumbnail image and/or a video frame, based on the user request and using a ranking algorithm, and event metadata associated with the preview image is received. Also for each event, a video segment associated with that event is received, and the preview image and/or the video segment is caused to be displayed to a user associated with the user request.

Systems and methods for trajectory-based object search using edge computing

A computer vision processor of a camera generates hyperzooms for persons or vehicles from image frames captured by the camera. The hyperzooms include a first hyperzoom associated with the persons or vehicles. The computer vision processor tracks traffic patterns of the persons or vehicles while obviating network usage by the camera by predicting positions of the persons or vehicles using a Kalman Filter from the first hyperzoom. The persons or vehicles are detected in the second hyperzoom. The positions of the persons or vehicles are updated based on detecting the persons or vehicles in the second hyperzoom. The first hyperzoom is removed from the camera. Tracks of the persons or vehicles are generated based on the updated positions. The second hyperzoom is removed from the camera. Track metadata is generated from the tracks for storing in a key-value database located on a non-transitory computer-readable storage medium of the camera.

People and vehicle analytics on the edge

A method for performing detection-based object searches includes receiving a user request indicating a region of interest, a timeframe of interest, or an object of interest. A signal is sent to cause execution of a query to identify object detections based on the user request. A signal representing at least one event identified in response to the query is received. For each event from the at least one event, a thumbnail image is identified based on the user request and using a ranking algorithm. A video frame identified based on the thumbnail image is received, and a video segment associated with the video frame is retrieved, A preview image clip that includes the video frame and the video segment is generated and displayed to a user associated with the user request.

Methods and apparatus for detection-based object search using edge computing

A computer vision processor of a camera extracts attributes of persons or vehicles from hyperzooms generated from image frames. The hyperzooms represent traffic patterns. The extracting is performed using a feature extractor of an on-camera convolutional neural network (CNN) including an inverted residual structure. The attributes include at least colors of clothing of the persons or colors of the vehicles. Mobile semantic segmentation models of the CNN are generated using the hyperzooms and the attributes. Attribute analytics are generated by executing the mobile semantic segmentation models while obviating network usage by the camera. The attribute analytics are stored in a key-value database located on a memory card of the camera. A query is received from the server instance specifying one or more of the attributes. The attribute analytics are filtered using the one or more of the attributes to obtain a portion of the traffic patterns.

Hyperzoom attribute analytics on the edge

A method for generating cropped images depicting objects captured in video data includes receiving, at a processor of a video camera system, a video stream including a series of video frames depicting an object. A classification for the object is generated, and an occurrence of the object being detected is identified in an additional video frame(s) from the series of video frames. A motion associated with the object is calculated based on the classification and the additional video frame(s). At least one image that depicts the object and includes a cropped portion of a video frame from the series of video frames is generated, along with an associated set of at least one quality score, in response to calculating the motion. The method also includes causing transmission of the at least one image to at least one remote compute device based on the set of at least one quality score.

Methods and apparatus for generating images of objects detected in video camera data

A method and related system operations include obtaining a video stream with an image sensor of a camera device, detecting a plurality of target objects by executing a neural network model based on the video stream with a vision processor unit of the camera device. The method also includes generating a plurality of bounding boxes, determining a plurality of character sequences by, for each respective bounding box of the plurality of bounding boxes, performing a set of optical character recognition (OCR) operations to determine a respective character sequence of the plurality of character sequences. The method also includes updating a plurality of tracklets to indicate the plurality of bounding boxes and storing the plurality of tracklets in association with the plurality of character sequences in a memory of the camera device.

Automatic multi-plate recognition

Automatic license plate recognition occurs when a light sensor that continually captures video detects motion as a vehicle is driven through a gate. The light sensor detects the vehicle and license plate in the video stream captured by the light sensor. An algorithm associated with the video stream of the light sensor is trained to detect license plates. The light sensor starts executing the recognition algorithm when it detects motion. Recognition of characters in the license plate is based upon an aggregation of several captured video frames in which a license plate is detected.

Automatic license plate recognition

Growing community of inventors

Belmont, CA, United States of America

Yi Xu