# US 12,602,795 B2 · False positive reduction of location specific event classification > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,602,795 B2 - Document ID: US12602795B2 - Status: Granted - Kind: B2 - Application number: 18/372,774 - Filed: 2023-09-26 - Published: 2025-03-27 - Granted: 2026-04-14 - Earliest claimed: 2023-09-26 - Assignees: Viisights Solutions Ltd. - Applicants: Viisights Solutions Ltd. - Inventors: Simon Polak; Shiri Gordon; Menashe Rothschild; Asaf Birenzvieg - Web page: https://patentgrep.com/patent/US12602795B2 ## CPC - G06T 7/0002: {Inspection of images, e.g. flaw detection} - G06T 7/246: using feature-based methods, e.g. the tracking of corners or segments - G06N 20/00: Machine learning - G06V 10/82: using neural networks - G06V 20/52: Surveillance or monitoring of activities, e.g. for recognising suspicious objects (recognising microscopic objects G06V20/69) - G06T 2207/20081: Training; Learning - G06T 2207/20084: Artificial neural networks [ANN] - G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING ## Abstract A method of training site-specific filter classifiers, comprising receiving a plurality of on-site feature vectors created by one or more classifiers deployed in a certain site to classify one or more predefined anomaly events based on imagery data captured in the certain site, the classifier(s) is trained using a plurality of non-site-specific samples, selecting a plurality of false positive feature vectors from a subset of the plurality of on-site feature vectors having a score exceeding a certain threshold, receiving a plurality of true positive feature vectors extracted from positive samples reflecting the predefined anomaly event(s),training one or more site-specific filter classifiers using the false positive feature vectors and the true positive feature vectors, and providing the trained site-specific filter classifiers for filtering out one or more false positive detections of the anomaly event(s) detected in the certain site by the classifier(s). ## Claims ### Claim 1 A method of training site-specific filter classifiers, comprising: using at least one processor for: receiving a plurality of on-site feature vectors created by at least one classifier deployed in a certain site to classify at least one predefined anomaly event based on imagery data captured in the certain site, the at least one classifier is trained using a plurality of non-site-specific samples; selecting a plurality of false positive feature vectors from a subset of the plurality of on-site feature vectors having a score exceeding a certain threshold; receiving a plurality of true positive feature vectors extracted from positive samples reflecting the at least one predefined anomaly event; training at least one site-specific filter classifier using the plurality of false positive feature vectors and the plurality of true positive feature vectors; and providing the at least one site-specific filter classifier for filtering out at least one false positive detection of the at least one anomaly event detected in the certain site by the at least one classifier. ### Claim 2 (depends on 1) The method of claim 1, wherein: each of the plurality of on-site feature vectors comprises a plurality of features extracted from a corresponding one of a plurality of on-site image samples captured by at least one image sensor deployed to monitor at least part of the certain site, and each of the plurality of true positive feature vectors comprises a plurality of features extracted from a respective one of the plurality of non-site-specific samples. ### Claim 3 (depends on 1) The method of claim 1, wherein the at least one site-specific filter classifier is integrated in the at least one classifier. ### Claim 4 (depends on 1) The method of claim 1, wherein the at least one site-specific filter classifier is independent of the at least one classifier. ### Claim 5 (depends on 1) The method of claim 1, wherein the at least one site-specific filter classifier comprises a plurality of site-specific filter classifiers each trained to filter out false positive detection of the at least one predefined anomaly event under different environmental conditions. ### Claim 6 (depends on 1) The method of claim 1, wherein a number of the plurality of false positive feature vectors exceeds a predefined number. ### Claim 7 (depends on 1) The method of claim 1, wherein the plurality of false positive feature vectors are extracted from a plurality of on-site samples captured during a time period exceeding a predefined time period. ### Claim 8 (depends on 1) The method of claim 1, wherein the training comprises: using a training subset to train the at least one site-specific filter classifier, the training subset comprises a first group of false positive feature vectors selected from the plurality of false positive feature vectors and a first group of true positive feature vectors selected from the plurality of true positive feature vectors, and using a validation subset to set a filter threshold for the at least one site-specific filter classifier such that each false positive feature vector having a filter classification score, computed by the at least one site-specific filter classifier, which does not exceed the filter threshold is filtered out, the validation subset comprises a second group of false positive feature vectors selected from the plurality of false positive feature vectors and a second group of true positive feature vectors selected from the plurality of true positive feature vectors. ### Claim 9 (depends on 1) The method of claim 1, wherein the training of the at least one site-specific filter classifier employs supervised learning wherein the false positive feature vectors having an exceeding score are annotated automatically. ### Claim 10 (depends on 1) The method of claim 1, further comprising further training the at least one site-specific filter classifier online, post deployment, without interfering with detection operation of the at least one classifier. ### Claim 11 (depends on 1) The method of claim 1, further comprising training the at least one site-specific filter classifier in at least one additional training iteration, wherein in each iteration the at least one site-specific filter classifier is further trained using a respective set of false positive feature vectors classified by the at least one classifier as positive in a previous iteration. ### Claim 12 (depends on 1) The method of claim 1, further comprising: identifying at least one uncertain on-site sample classified as false positive which is estimated to potentially be true positive, analyzing the at least one uncertain on-site sample to determine whether it is true positive or false positive, annotating the feature vector extracted from the at least one uncertain on-site sample as positive or negative according to the determination, and using the annotated feature vector for further training the at least one site-specific filter classifier. ### Claim 13 (depends on 12) The method of claim 12, further comprising applying at least one image processing algorithm to remove privacy features from the at least one uncertain on-site sample. ### Claim 14 (depends on 12) The method of claim 12, wherein the at least one on-site sample is manually analyzed and annotated as positive or negative. ### Claim 15 (depends on 1) The method of claim 1, further comprising deploying at least one updated classifier to replace the least one classifier, and training the at least one site-specific filter classifier using at least one on-site feature vector extracted from at least one stored on-site sample captured prior to the update and classified by the at least one updated classifier. ### Claim 16 (depends on 1) The method of claim 1, wherein the at least one site-specific filter classifier comprises at least one machine learning model. ### Claim 17 A system for training an on-site filter classifier, comprising: a memory storing a code; and at least one processor configured to execute the code, the code comprising: code instructors to receive a plurality of on-site feature vectors created by at least one classifier deployed in a certain site to classify at least one predefined anomaly event based on imagery data captured in the certain site, the at least one classifier is trained using a plurality of non-site-specific samples; code instructors to select a plurality of false positive feature vectors extracted from a subset of the plurality of on-site feature vectors having a score exceeding a certain threshold; code instructors to receive a plurality of true positive feature vectors extracted from positive samples reflecting the at least one predefined anomaly event; code instructors to train at least one site-specific filter classifier using the plurality of false positive feature vectors and the plurality of true positive feature vectors; and code instructors to provide the at least one site-specific filter classifier for filtering out at least one false positive detection of the at least one anomaly event detected in the certain site by the at least one classifier. ### Claim 18 A method of reducing false positive classification events using site-specific filter classifiers, comprising: using at least one processor for: receiving at least one on-site sample classified as positive by at least one classifier deployed in a certain site to detect at least one predefined anomaly event based on imagery data depicting at least part of the certain site; applying at least one site-specific filter classifier to compute a filter classification score for an on-site feature vector extracted from the at least one on-site sample; estimating whether the at least one on-site sample is true positive or false positive based on comparison of the filter classification score to a filter threshold; filtering out the at least one on-site sample in case the filter classification score does not exceed the filter threshold; and outputting at least one detection indication in case the filter classification score exceeds the filter threshold indicative of detection of the at least one predefined anomaly event; wherein the at least one classifier is trained using a plurality of non-site-specific samples; wherein the at least one site-specific filter classifier is trained using a plurality of false positive feature vectors and a plurality of true positive feature vectors, the plurality of false positive feature vectors are extracted from a subset of a plurality of on-site samples having a score, computed by the at least one classifier, exceeding a certain threshold, and the plurality of true positive feature vectors extracted from a subset of the plurality of non-site-specific samples reflecting the at least one predefined anomaly event.