Time-lapse video generation centered on events / objects of interest with the help of neural network input on camera devices
By introducing neural networks on camera equipment, detecting events of interest and dynamically adjusting the frame rate, the problem of fixed frame rate in the intermediate interval shooting video mode in the prior art is solved, and the generation of intelligent interval shooting videos is realized and efficient data transmission is improved. Video browsing efficiency and equipment performance.
Patent Information
- Application Number
- CN202010836914.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-08-19
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2040-08-19
AI Technical Summary
In the interval video mode, the cameras implemented by existing Internet-connected clouds are difficult to effectively utilize the captured video data, making it difficult for users to quickly find important details, and the frame rate is fixed and cannot adapt to changes in events of interest.
Introduce neural networks on camera equipment to realize dynamic adjustment of event/object detection and video frame rate, generate intelligent interval shooting video based on events of interest, detect objects and adjust frame rate through convolutional neural network, generate encoded video frames , transmit to cloud storage.
It realizes real-time analysis of video content on edge devices, dynamically adjusts frame rate, ensures the normal display speed of content of interest, improves video browsing efficiency, avoids excessive consumption of cloud server resources, and provides a more energy-saving solution.
Smart Images

Figure CN114079820B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates generally to computer vision, and more particularly to methods and / or apparatus for enabling timelapse video generation centered around an event / object of interest on a camera device with the aid of neural network input. Background Art
[0002] The time-lapse video mode for conventional Internet-connected / cloud-enabled cameras is typically implemented by software operating on a cloud server. The software relies on the distributed processing resources (i.e., scalable computing) of the cloud server. The time-lapse video clips are displayed at a fixed frame rate to achieve a fast-forward effect. For example, video frames are selected at fixed intervals to create a time-lapse video (i.e., every 30th frame from a thirty frames per second video is selected to create a time-lapse video).
[0003] Internet-connected / cloud-enabled cameras encode video data and then transmit the encoded video stream to a cloud server. Many extra CPU cycles are used on the cloud server to decode a regular video stream (i.e., compressed video using AVC encoding or HEVC encoding), extract frames at fixed frame intervals, and then transcode those frames into timelapse video. At the same time, even with timelapse video, it is difficult for users to find important details. Because timelapse video always uses a fixed frame rate and does not use all of the captured video data, video frames at normal display speed are not available for the duration of the content of interest to the user in the captured video. For example, when a security camera observes a potential event of interest (e.g., a pedestrian), the timelapse will be the same as when the security camera does not observe any event of particular interest.
[0004] It would be desirable to enable time-lapse video generation centered around an event / object of interest with the aid of a neural network input on a camera device. Summary of the invention
[0005] The present invention relates to a device, which includes an interface and a processor. The interface can be configured to receive pixel data generated by a capture device. The processor can be configured to perform the following operations: receiving pixel data from the interface; generating a video frame in response to the pixel data; performing computer vision operations on the video frame to detect an object; performing classification of the detected object based on the characteristics of the object; determining whether the classification of the object corresponds to a user-defined event; and generating an encoded video frame based on the video frame. The encoded video frame can be transmitted to a cloud storage service. The encoded video frame may include a first sample of the video frame and a second sample of the video frame, the first sample being selected at a first rate when the user-defined event is not detected, and the second sample being selected at a second rate when the user-defined event is detected. The second rate may be greater than the first rate. BRIEF DESCRIPTION OF THE DRAWINGS
[0006] Embodiments of the present invention will become apparent from the following detailed description and appended claims and the accompanying drawings.
[0007] Figure 1 is a diagram illustrating an example context of the present invention.
[0008] Figure 2 is a diagram illustrating an example Internet-connected video camera implementing an example embodiment of the present invention.
[0009] Figure 3 is a block diagram illustrating components of an apparatus configured to provide event-centric time-lapse video with the aid of a neural network.
[0010] Figure 4 is a diagram showing interconnected cameras communicating with a cloud server and a video processing pipeline for generating time-lapse video.
[0011] Figure 5 is a diagram showing the smart interval shooting mode on the edge AI camera with CV analysis of all frames using events.
[0012] Figure 6 is a diagram showing the smart interval shooting mode on an edge AI camera with CV analysis of partial frames using events.
[0013] Figure 7 is a diagram illustrating event detection in captured video frames.
[0014] Figure 8 is a diagram showing an application for controlling preferences for time-lapse video operated on a smartphone. DETAILED DESCRIPTION
[0015] Embodiments of the present invention include providing time-lapse video generation centered on an event / object of interest with the aid of a neural network input on a camera device, and embodiments of the present invention may perform the following operations: (i) implement event detection on an edge device; (ii) implement video encoding on an edge device; (iii) generate smart time-lapse video with a varying frame rate based on the detected event / object; (iv) detect objects / events using a convolutional neural network implemented on a processor; (v) adjust the time-lapse frame rate to capture all video frames when an event / object is detected; (vi) perform facial recognition and / or object classification; (vii) upload the encoded smart time-lapse video to a cloud storage server; (viii) use a smartphone app to implement configuration of parameters for event / object detection and / or the frame rate of the time-lapse video; and / or (ix) be implemented as one or more integrated circuits.
[0016] Embodiments of the present invention may be configured to provide a smart interval shooting video mode. The interval shooting video mode may be a smart interval shooting video mode that automatically adjusts the video display speed based on the content captured in the video. Embodiments of the present invention may be configured to generate encoded smart interval shooting videos. Smart interval shooting videos may be generated on a cloud service (e.g., using scalable computing). Smart interval shooting videos may be generated on an edge device (e.g., an artificial intelligence (AI) camera).
[0017] Edge AI cameras can be configured to implement artificial intelligence (AI) technology. Edge AI cameras use AI technology that can be more powerful in many ways than using cloud servers (e.g., by providing relevant data to users) and is a more energy-efficient solution. Edge AI cameras can be configured to execute computer-readable instructions on the device (e.g., without relying on external processing resources) to analyze video content frame by frame. Based on the analysis, the content can be tagged with metadata information. Metadata can be used to select video frames for time-lapse video. For example, video frames can be classified by being marked as not having an object / event of interest and having an object / event of interest. Computer vision (CV) operations can determine whether there is an object / event of interest.
[0018] When there is no CV event (or type of object) of interest for a video frame for a duration of N seconds (N=60 / 120 / 240 / …), the edge AI camera may select one of the video frames from the no-event duration. The selected frame may be used for video encoding (e.g., a video encoder built into the edge AI camera device). Selecting a video frame from the no-event duration N for encoding may be repeated for each N-second duration in which no event is detected. Selecting a video frame from the no-event duration N for encoding may result in an encoded output that provides a compressed portion of the captured video (e.g., thereby efficiently fast-forwarding through a “meaningless content” portion of the captured video at a high display speed).
[0019] When there is a detected CV event (or type of object) of interest for a video frame within a duration of M seconds (M=5 / 15 / 30 / …), the edge AI camera can adjust the number and / or selection rate of video frames within the event duration M. In an example, the object and / or event of interest can be a person detected, a car detected, an animal detected, an action, a specific face detected, etc. The selection of video frames can be adjusted to select all video frames from the M-second duration of the event (e.g., for a 2-minute event captured at 60 frames per second, all 7200 frames can be selected). The selection of video frames can be adjusted to select video frames at a higher rate than in the event-free duration within the M-second duration of the event (e.g., more frames are selected, but not all frames), for example, for a 2-minute event captured at 60 frames per second, the selected rate can be changed to 30 frames, and every other frame can be selected, resulting in the selection of 3600 video frames. The selected video frames can be encoded (e.g., using on-device video encoding of the AI edge camera). Selecting video frames for encoding at a higher rate within the event duration M can result in an encoded output that provides a portion of the smart time-lapse video having a normal display speed for "meaningful content", or a portion of the smart time-lapse video having a slightly compressed display speed for "meaningful content" (but not as high as the speed for "meaningless content").
[0020] With the smart timelapse video mode implemented on the edge camera, a user can quickly browse video content captured over a long period of time (e.g., days / weeks / months), and the user can be confident that CV events of interest will not be missed. CV events of interest may include detecting people (e.g., known people) through face detection and / or face recognition, detecting car license plates (e.g., using a license plate reader), detecting pets (e.g., known animals) using animal / pet recognition, detecting motion (e.g., any motion detected above a predefined threshold), etc. The types of detected events and / or objects that may be considered events of interest may vary according to the design criteria of a particular implementation.
[0021] Embodiments of the present invention may enable a user to specify the types of objects / events that may be considered to be of interest. In one example, an app operating on a smartphone may be configured to adjust settings for an edge AI camera. In another example, the edge AI camera may be configured to provide a web interface to enable a user to remotely select objects / events that are considered to be events of interest. In yet another example, the edge AI camera may be connected to a cloud server, and the user may use the web interface to adjust settings stored on the cloud server, which may then be sent to the edge AI camera to control the types of timelapse objects of interest.
[0022] The duration of the timelapse can be configured as a regular interval, and can also be configured as the time since the user last received a timelapse video. For example, if the user has missed 20 event notifications, at the moment the user interacts with the app (e.g., swiping on a smartphone) to view the event, a timelapse focused on the event can be presented to the user for viewing. Although embodiments of the present invention can generally be performed locally on an edge AI camera (e.g., thereby implementing a processor configured to implement a convolutional neural network and / or video encoding), embodiments of the present invention can be performed using software on a cloud server to achieve a similar effect.
[0023] refer to Figure 1, a diagram illustrating an example context of the present invention is shown. A residence 50 and a vehicle 52 are shown. A camera system 100a-100n is shown. Each of the cameras 100a-100n can be configured to generate smart time-lapse videos. For example, each of the cameras 100a-100n can be configured to operate independently of each other. Each of the cameras 100a-100n can capture video and generate smart time-lapse videos. In one example, the corresponding smart time-lapse videos can be uploaded to a cloud storage service. In another example, the corresponding smart time-lapse videos can be stored locally (e.g., stored on a microSD card, stored to a local network attached storage device, etc.).
[0024] Each of the cameras 100a-100n can be configured to detect different or the same events / objects that can be considered to be of interest. For example, the camera system 100b can capture an area near the entrance of the residence 50. For the entrance of the residence 50, the object / event of interest can be detecting a person. The camera system 100b can be configured to analyze the video frames to detect a person, and when a person is detected, the smart interval shooting video can be slowed down (e.g., selecting a video frame for encoding at a higher frame rate). In another example, the camera system 100d can capture an area near the vehicle 52. For the vehicle 52, the object / event of interest can be detecting other vehicles and pedestrians. The camera system 100b can be configured to analyze the video frames to detect vehicles (or road signs) and people, and when a vehicle or pedestrian is detected, the smart interval shooting video can be slowed down.
[0025] refer to Figure 2 , a diagram illustrating an example Internet-connected camera implementing an example embodiment of the present invention is shown. Camera systems 100a-100n are shown. Each camera device 100a-100n can have a different style and / or use case. For example, camera 100a can be an action camera, camera 100b can be a ceiling-mounted security camera, camera 100n can be a network camera, etc. Other types of cameras (e.g., residential security cameras, battery-powered cameras, doorbell cameras, etc.) can be implemented. The design / style of cameras 100a-100n can vary depending on the design criteria of a particular implementation.
[0026] The camera systems 100a-100n may include a processor 102 and a capture device 104. The processor 102 may be configured to implement a convolutional neural network (CNN). The processor 102 may be configured to implement a video encoder. The processor 102 may generate smart time-lapse video. The capture device 104 may be configured to capture pixel data that may be used to generate video frames.
[0027] refer to Figure 3 , a block diagram illustrating components of an apparatus configured to provide event-centric time-lapse video with the aid of a neural network is shown.
[0028] The processor 102 may be configured to generate encoded video frames based on pixel data received from the capture devices 104a-104n. The processor 102 may be configured to select video frames to be encoded for smart time-lapse video. Video frames may be selected in response to objects / events detected by the CNN module 190b. The scheduler circuit 190a may load a directed acyclic graph (DAG) including parameters for detecting various types of objects / events. For example, different DAGs may be loaded in response to the type of event of interest selected by the user.
[0029] refer to Figure 4 , a diagram illustrating interconnected cameras communicating with a cloud server and a video processing pipeline for generating time-lapse video is shown. The edge AI camera can be configured to communicate with a remote cloud service and / or a user device (e.g., a smartphone). The user can select settings using an app on a smartphone 52b-52n. Preferences can be stored in the cloud service as event settings. The cloud service can provide neural network parameters (e.g., feature sets) to the processor 102. The processor 102 can receive feature sets to detect objects / events. The edge AI camera can be configured to upload encoded video frames (e.g., smart time-lapse videos) to the cloud service. The encoded video frames can be stored in a large-capacity video storage device of the cloud service. The user can use a smartphone to access the encoded video frames. The smartphone can connect to the large-capacity video storage device and download the smart time-lapse video. The cloud service can further provide notifications to the smartphone based on the content in the metadata tag of the smart time-lapse video.
[0030] The video processing pipeline of the processor 102 is shown. The processor 102 can generate raw video frames in response to pixel data captured by the capture device 104. The raw video frames can be analyzed by the CNN module. The CNN module can include an AI model (e.g., DAG). The CNN module can detect objects of interest based on the user's preferences (e.g., detecting one or all of a person, a vehicle, an animal, movement, a sound, etc. as an event).
[0031] When an event is detected, the CNN module can mark the video frame. The video frame marked with the event can be provided to the video encoding pipeline. The video frame with the detected event can be encoded at an adjusted frame rate (e.g., the full video or a higher frame rate than when the event is not detected).
[0032] The CNN module may determine which frames do not have detected events (e.g., video frames that are not of interest). Among the video frames that do not have events, the processor 102 may select a video frame for each duration of length N. For example, the processor 102 may select a video frame from the duration N in which the video is captured at 60 fps (e.g., select a frame from 60 frames captured in one second).
[0033] The unselected video frames may not be included in the smart time-lapse video stream. In one example, the unselected video frames may be discarded (e.g., not stored). In another example, the unselected video frames may be used in a recording (e.g., a recording of the complete video stream that may be stored locally on the camera device 100, stored on a microSD card, etc.). The selected video frames may be presented to a video encoding pipeline for encoding into a smart time-lapse video.
[0034] The video encoding pipeline can be configured to encode video frames in real time (e.g., HEVC, AVC, H264, etc.). The video encoding pipeline can insert selected video frames into the encoded video in sequence based on the time of capture. For example, each video frame can have a timestamp, and the video frames can be inserted into the encoded video in the order of the timestamps (e.g., the earliest video frame first to create a chronologically ordered video). The video encoding pipeline can generate an encoded video with video frames with events (displayed at normal speed) and video frames of non-events (displayed at a faster speed because fewer frames are selected).
[0035] The video encoding pipeline can be configured to apply various effects to selected video frames. For example, the video encoding pipeline can be configured to crop selected video frames, dewarp selected video frames, apply colorization (or remove color) based on selected video frames, apply blur to segments of video frames, etc. In an example, blur effects can be applied to portions of selected video frames to hide and / or censor specific objects (e.g., to hide the identity of known people for privacy reasons, to hide nude parts to prevent nude content from being uploaded, etc.).
[0036] Processor 102 may be configured to perform computer vision operations before video frames are encoded. Performing computer vision operations before video frames are encoded may enable computer vision operations to be performed without the cycle cost of first decoding the video frames. For example, if computer vision is performed by a cloud service, the encoded video frames are sent to a cloud server, and the cloud server wastes CPU cycles and power by first decoding the video frames, then performing computer vision operations, and then transcoding the time-lapse video.
[0037] The processor 102 may be configured to perform video frame selection for smart time-lapse video all locally at the edge AI camera 100. The processor 102 may implement a CNN module. The CNN module may be a hardware component (or multiple hardware components working together) of the processor 102 that implements a CNN-based detection technique (or other types of detection techniques) that all operate on the processor 102 to perform computer vision operations such as object detection, object tracking, object classification, etc. The hardware-accelerated CNN module may enable object / event detection to be performed on the camera system 100 (e.g., without relying on the computing resources of a cloud server to perform object detection). In some embodiments, the processor may use software acceleration for computer vision and / or a combination of hardware-accelerated computer vision and software-accelerated computer vision. Using computer vision operations running on a hardware-based CNN engine and / or a software-based CNN engine, the processor 102 implemented in the edge device 100 may be able to detect one or more types of events and apply event markers to captured video frames.
[0038] refer to Figure 5 , a diagram illustrating a smart interval shooting mode on an edge AI camera with CV analysis of all frames using events is shown. A captured video stream including a stream of video frames (e.g., a group of pictures (GOP)) is shown. The CNN module can perform computer vision to determine which video frames include CV events. The first four groups of duration N of video frames may not include CV events. One of the video frames from each of the first four groups of duration N can be selected by the processor 102 for smart interval shooting video (e.g., four non-event video frames are selected).
[0039] After four non-event groups, a CV event may be detected (e.g., of duration M). For a complete smart time-lapse video, all video frames with CV events within duration M may be selected for the smart time-lapse video stream. In the example shown, there may be six video frames within duration M. All six video frames with CV events may be selected for the smart time-lapse video.
[0040] After the duration M of the CV event, two other groups of duration N of video frames that do not include the CV event may be detected. The frame rate for video selection may be readjusted back to the default rate. For example, a video frame may be selected for each of the N groups for the duration of the next two non-events of the video frame. Although the first video frame is shown as being selected from the non-CV event group of video frames, any one of the video frames may be selected from each of the non-CV event groups according to the design criteria of a particular implementation.
[0041] Smart interval shooting video can be played back at a normal frame rate. Since there are fewer video frames to be played back from the non-CV portion of the video frame, the playback of the non-CV portion of the smart interval shooting video stream may appear to be played back at a higher speed than real-time playback. For example, if 60 non-CV event groups of video frames have been detected, a frame can be selected from each frame for the smart interval shooting video. Playing back the smart interval shooting video at 60fps can cause the non-CV event group to be played back in a compressed form of one second. Since all the video frames of the CV event have been selected, the playback of the portion of the smart interval shooting video stream can be played back at a normal speed (e.g., in real time).
[0042] The processor 102 can generate a smart time-lapse video based on the detected event. The CNN module can detect events and mark video frames corresponding to specific events. In one example, a video frame (e.g., frame N) can correspond to event A and event B, and the next one hundred video frames after frame N (e.g., frame N+100) can correspond only to event B. The processor 102 can be configured to combine the selected frames in the following manner: select some video frames from a video without an event (e.g., typically only select I-frames), and then select all frames with event A markings. Then, the encoding device can create a smart time-lapse video (e.g., compressed video) including event A (e.g., by finding frame N and any adjacent frames with event A markings that can be selected). Similarly, different smart time-lapse videos (e.g., separate video files) can be generated for event B (e.g., by finding video frame N and the next frame N+100 with event B markings).
[0043] refer to Figure 6 , a diagram illustrating a smart timelapse mode on an edge AI camera with CV analysis of partial frames using events. In some embodiments, not all video frames of a CV event may be selected. Not selecting all video frames of a CV event may result in the smart timelapse video providing a compressed view of the CV event that still provides more video frames (e.g., more details, slower playback speed) than a compressed view of selected non-CV event video frames.
[0044] In the example shown, four non-CV events of duration N can select one video frame for the smart interval shooting video stream. After four non-event groups, a CV event (e.g., duration M) can be detected. For a partial smart interval shooting video, not all video frames with CV events within duration M can be selected for the smart interval shooting video stream. In the example shown, the frame rate for selection can be half of the video capture frame rate. In the example shown, there can be six video frames within duration M. Three of the six video frames with CV events can be selected for the partial smart interval shooting video.
[0045] exist Figure 5 and Figure 6 , the video frames selected by the processor 102 to be encoded for the smart time-lapse video are shown as larger than the frames that may not be selected. For illustrative purposes, the sizes of the selected video frames and / or the unselected video frames are shown as different (e.g., a larger size indicates that the video frame has been selected, and a smaller size indicates that the video frame has not been selected). In general, the selected video frames may not necessarily be video frames of a particular type (e.g., the selected video frames may be any of an I-frame, a B-frame, a P-frame, etc.). For example, all video frames in the "CV event duration" may be part of the same coded GOP as the video frames in the "no CV event duration", which may enable the selected video frames to be any type of video frames. In the example, the selected frames may not all need to be I-frames, because when the video frames are captured and the processor 102 is performing encoding, object / event detection may be performed in parallel by the CNN module (e.g., the computer vision operations by the CNN module may not have to be completed as soon as a new video frame is captured). Computer vision operations may be performed, and then if an event is detected, the entire video GOP may be marked as an event (e.g., event A) without changing the video GOP for the event. Processor 102 can decode the entire GOP and select some of the frames using the CNN module. For example, for a portion of the smart time-lapse video (in Figure 6 ), for better quality, the CNN module may decide to search only some of the video frames in a GOP.
[0046] refer to Figure 7 , a diagram illustrating event detection in a captured video frame is shown. An example video frame is shown. The example video frame may include pixel data provided to the CNN module. The CNN module may be configured to analyze the pixel data of the video frame to determine whether an event / object of interest exists. The type of object / event detected by the CNN module may be defined by a feature set selected by a user and / or loaded by the processor 102.
[0047] An example video frame may be a captured video of the front of residence 50 (e.g., a view from an angle looking down the driveway from the front door). An example video frame may include a person signing for a package delivered by a delivery person. A delivery truck may be parked on the road. Trees and nearby residences are also shown.
[0048] A dashed box representing a bounding box of an object detected by the CNN module is shown. The detected objects may be a face of a delivery person, a face of a person (e.g., a homeowner), and a delivery vehicle. In an example, the settings for processor 102 may define objects of interest as people and vehicles. For example, trees and homes may not be objects of interest.
[0049] Since the example video frame includes at least one object of interest, the CNN module can mark the video frame as an "event" video frame. If the smart time-lapse video includes all video frames within the duration of the event, the example video frame can be encoded and transcoded as a smart time-lapse video frame.
[0050] refer to Figure 8 , a diagram illustrating an application operating on a smartphone for controlling preferences for timelapse video is shown. An app interface is shown. The smartphone can be configured to connect directly to the edge AI device. The smartphone can be configured to connect to a cloud service to access user preferences stored in the cloud. The edge AI device can be configured to operate according to settings retrieved from the cloud.
[0051] The App can enable the user to choose whether to encode the complete smart timelapse video stream (e.g., select all video frames with detected events / objects) or a portion of the timelapse video stream. If the partial smart timelapse video stream option is selected, the user can further define the selected frame rate for the video frames with detected events. In the example shown, 60fps, 30fps, and 15fps are shown as available options. The available options for the frame rate of the partial smart timelapse video can vary depending on the design criteria of the specific implementation. Selecting a lower frame rate can result in a more compressed version of the detected event being inserted into the smart timelapse video stream.
[0052] The app can enable the user to select the type of object and / or event to be detected. For example, an event can be the detection of a person, vehicle, animal (e.g., using computer vision), detecting an action, etc. An event can be an audio event (e.g., detecting sounds that may be louder than a predefined sound level, a specific type of sound signature, etc.). The app can further enable the user to watch a live stream (e.g., displaying video data currently being captured by the edge AI device). The app can provide a notification when an event is detected by the CNN module.
[0053] The terms "may" and "generally" when used in conjunction with "is" and a verb herein are meant to convey such an intention: the description is exemplary and is considered broad enough to cover both the specific examples presented in this disclosure and alternative examples that can be derived based on this disclosure. The terms "may" and "generally" as used herein should not be interpreted as necessarily implying the desirability or possibility of omitting the corresponding element.
[0054] While the invention has been particularly shown and described with reference to embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made without departing from the scope of the invention.
Claims
1. A device comprising: an interface configured to receive pixel data generated by a capture device; as well as a processor configured to: (i) receive the pixel data from the interface; (ii) generate a video frame in response to the pixel data; (iii) perform computer vision operations on the video frame using a neural network to detect objects / events, wherein the processor receives neural network parameters from a cloud storage service to detect user-defined objects / events; (iv) performing classification of the detected object / event based on characteristics of the event / object to tag video frames with metadata for the user-defined object / event; (v) determining whether the classification of the event / object corresponds to a user-defined object / event; and (vi) generating an encoded video frame from the video stream of video frames, wherein: (a) the encoded video frame is transmitted to the cloud storage service; (b) the video stream comprises a plurality of video frames of a non-event group and at least one video frame of an event group, wherein the video frames of the non-event group are not marked with a user-defined event and each video frame has a first duration, and the video frames of the event group are marked with a user-defined event and each video frame has a second duration different from the first duration; (c) the encoded video frame comprises: (i) a frame selected from the non-event group of video frames when the user-defined object / event is not detected within the first duration, and (ii) the video frame selected from the event group of video frames when the user-defined object / event is detected for a video frame within the second duration; (d) The video frames selected from the encoded video frames are arranged according to the timestamp of each video frame.
2. The device according to claim 1, wherein: The rate for selecting video frames from the video frames of the event group is the same as the frame rate of the video frames.
3. The device according to claim 1, wherein: A rate for selecting video frames from the video frames of the event group is less than a frame rate of the video frames.
4. The device according to claim 1, wherein: The computer vision operations are performed on an integrated circuit.
5. The device according to claim 1, wherein: (i) the apparatus and the capture device are implemented on an edge device, and (ii) the edge device communicates with the cloud storage service.
6. The device according to claim 1, wherein: The selected video frame includes any one of an I-frame, a B-frame or a P-frame.
