Video Head Tracking and Redaction for Missed Face Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The manual redaction of bystander faces in police body camera videos is time-consuming and inefficient, and conventional head detection algorithms fail to detect faces when individuals are not facing the camera directly, potentially exposing innocent bystanders to privacy risks.
Innovation Solution
A two-layer head detection system using a pre-trained neural network (e.g., YOLOv3) for initial detection, followed by an optical classification engine (e.g., SVM or dlib correlation tracking) for reanalysis of frames where heads may be missed, ensuring comprehensive face detection and redaction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual redaction is used, then accuracy of face detection is improved, but time consumption and labor resources increase significantly
Solution Approach 1:
The patent replaces manual mechanical redaction processes with an automated computer vision system using neural networks (YOLOv3) to detect faces and optical flow algorithms to track movements, eliminating the need for manual frame-by-frame inspection while maintaining detection accuracy
Solution Approach 2:
The system creates a digital copy of the video footage and processes only this copy through automated detection algorithms, allowing the original video to remain unchanged while enabling efficient redaction of the processed version without time-consuming manual intervention
2Productivity
If conventional head detection algorithms are used, then processing speed is improved, but detection accuracy deteriorates when individuals are not facing the camera directly
Solution Approach 1:
The patent implements a feedback mechanism where the optical flow algorithm continuously tracks the movement and orientation of detected faces, adjusting the detection parameters in real-time to maintain accurate detection even when individuals turn their heads or move out of direct view
Solution Approach 2:
The system transitions from static face detection to dynamic tracking using optical flow, allowing the detection algorithm to adapt to changing orientations and movements of individuals in real-time, maintaining accuracy regardless of whether subjects are facing the camera directly
3Productivity
If automated redaction systems are implemented, then time efficiency is improved, but detection reliability worsens due to missed faces in challenging scenarios
Solution Approach 1:
The patent introduces optical flow as an intermediary mechanism between the initial neural network detection and the final redaction process, serving as a mediator that tracks face movements and orientations to identify additional faces that may have been missed by the primary detection system
Solution Approach 2:
The system performs preliminary detection using neural networks to identify potential faces, then uses optical flow analysis as a preliminary verification step to track and confirm face movements and orientations before finalizing the redaction, ensuring reliable detection in challenging scenarios
Data Source
AI summary
Disclosed are systems and methods to detect and track an object across frames of a video. One of the disclosed methods includes: detecting a first group of one or more objects, using a first neural network, in each frame of the video, wherein each detected head of the first group comprises a leading and a trailing edge; grouping the leading and trailing edges of the one or more objects into groups of leading edges and groups of trailing edges based at least on coordinates of the leading and trailing edges; generating a list of no-edge-detect frames by identifying frames of the video missing a group of leading edges or a group of trailing edges; analyzing the no-edge-detect frames in the list of no-edge-detect frames, using an optical image classification engine, to detect a second group of one or more objects in the no-edge-detect frames; and merging the first and second groups of one or more objects to form a merged list of detected objects in the video.