Research in Applications for Learning Machines (REALM) Consortium - PowerPoint PPT Presentation

AFRL, Rome Request for Information • The Air Force Research Laboratory, Information Directorate (AFRL/RI) is seeking information to better understand existing vendor offerings within the landscape of research and development (R&D) that could later drive the development of prototypes of Machine Learning (ML) enabled Operational Command & Control (C2) functions and assess their notional value1, 2 to Operational C2. • The Air Force is investigating the incorporation of Machine Learning capabilities into Air Force C2 applications. As such, it is interested in the identification of C2 applications that can benefit from the incorporation of these capabilities, an understanding of how these applications and operations can notionally benefit, and the algorithms, and necessary data that will be a part of these implementations. This RFI is requesting information to better understand those AF C2 applications that have incorporated ML, those that could incorporate ML in the future and the algorithms which support these advanced capabilities. The C2 applications should fall into one of the following categories: Operational C2 supporting the air tasking process, battle management supporting operations execution, tactical-level C2 supporting the end-user, and Multi Domain C2. 21

Army Research Lab • STRONG addresses a critical objective within a broader Army goal to enable effective integration of Artificial Intelligence / Machine Learning (AI/ML) in the battlefield. This program has been developed in coordination with other related ARL-funded collaborative efforts (see descriptions of ARL collaborative alliances at https://www.arl.army.mil/www/default.cfm?page=93) and shares a common vision of highly collaborative academia-industry-government partnerships; however, it will be executed with a program model different than previous ARL Collaborative Research/Technology Alliances. 22

Long Term Objectives of Research • Retrieve knowledge for multiple users’ changing needs and mission. Relate multi- modal data and dynamically update/build the knowledge base for users. Utilize users' queries to build knowledge on top of a relational database and cache appropriate data and queries to improve performance. Lean about Knowledge graphs from ISI research. • Integrate new streaming data with knowledge queries already used by mission. Complete the unfulfilled data needs for missions. Discover new knowledge that can benefit mission • Conduct research in learning machines to make this efficient at large scale • Research transfer learning, reinforcement learning, active learning and apply to NG large databases ( sensors, signals, text, phone calls, videos, images, voice) Some of these are long term objectives. Include efficient labeling, NLP Slide 23 • Make system practical and responsive and efficient by using systems, ML, and tools already available and used in industry

Data Management Data at rest. Segments of video as tuples in the DB. Feature analysis. Queries model the knowledge. Postgres Slide 24

Queries model the knowledge • Related queries build knowledge. Query • All queries can relate to one engine event. • Mission changes with the data stream. Queries -> Knowledge Slide 25

Queries model the knowledge • Alerts • Triggers Query • Cached queries engine Cache hot Cache recently used queries data Slide 26

Situational Knowledge Query Engine Architecture Slide 27

Scenario: Save Child Left Alone in Car in heat or cold • In 2019, 51 children died from heatstroke after being left in a hot vehicle, 2 in Indiana.* Context & User Mission Contextual Info. Propagation Situational Normal Day & Bad Finding an Unattended Information Regular Petrol Send to Appropriate User Child in Car SKOD forwarded to Good During an Earthquake & Finding an Unattended Appropriate Rescue Personnel Send to Appropriate User Child in Car User City Data 28 * https://injuryfacts.nsc.org/motor-vehicle/motor-vehicle-safety-issues/hotcars/

Scenario: Stop Suspected Person from Violence ATF Records Context: NY Police New Years needs to • Record of people buying guns and ammunitions in Evening Know an area Suspected BMV Records Person • Record of DUI Convictions GPS tracking • Headed to NYC times square crimemapping.com • Is involved in Assault / Census Records Disturbing the peace / Homicide / Vandalism • No Family Connection to NYC or close by 32

Urban Information System Scenarios Identify Unsafe Lane Changes Identify Jaywalking 33

City of Cambridge: Agents • Numerous agents with different missions in a city (i.e., Cambridge) – Cambridge police – University (Harvard, MIT) police – TRANSIT police – Cambridge public works – Citizens – FEMA ( Emergency personnel) – Homeland Security 34

Missions • Missions with various needs for information – MIT police (pedestrians in the middle of the road, unsafe lane changes, ”choke” points, Child left alone in parked car, purple Cadillac used by a bad guy identified …) – Cambridge public works (potholes, down or occluded street signs) – Citizens (crane or car illegally blocking the sidewalk in front of house) 35

SKOD Objectives SKOD Service SKOD Service • Retrieve knowledge needed by multiple users with changing needs based on Situational Awareness Data Repository Data Repository Learning Machine Engine Learning Machine Engine • Relate multi-modal data and update the knowledge for users All available data All available data Data Requests Data Requests Knowledge Discovery Engine Knowledge Discovery Engine Access Access Deep Learning Module Deep Learning Module Pattern DB Pattern DB • Integrate new streaming data with queries already used by mission Pattern Recognition Pattern Recognition New data item User 1 User 1 • Complete the unfulfilled data needs for missions based on the Situation Recommended data after Data Controller Data Controller Recommended processing and User Preference User Profiling User Profiling data for User 1 User 2 User 2 - - Preferences Preferences - - Roles Roles - - Context Context Objective 2: New data items are directed to interested users based Objective 1: Relevant data is efficiently passed to users based on on User Profiling. their requests 36

Datasets Collected for City of Cambridge • Video – 100+ hours of dashcam video collected at MIT – Raw video can be retrieved from MIT database at Cambridge • Split into chunks of 30 seconds • Metadata collected: geolocation and timestamp for each 30 seconds • Unstructured Text (Twitter data) – Collected ~200K tweets (Target ~ 1 million) – Automatic tweet parsing and recording system into Postgres in place • Structured data – Cambridge public datasets – Automatic weekly updates into Postgres in place • Data from drones and dashcams 37 37

Datasets Example • Tweets from Cambridge Police • A video that has a bicyclist without helmet on it 00:01 to 00:27 38

Future Datasets • Waymo Open Dataset – Sensor data • Synchronized lidar and camera data from 1,000 segments (20s each) – Labeled data • Labels for 4 object classes - Vehicles, Pedestrians, Cyclists, Signs • Yelp Dataset – Reviews – Businesses – Pictures – Metropolitan Areas • News Articles – https://www.cambridgema.gov/news?page=2&ResultsPerPage=10 – Google News https://waymo.com/open/; https://www.yelp.com/dataset 39

Feature Extraction Architecture ML, NLP Index Constructor Multimodal Streaming Data type Processors Data NLP (Text) Data Sources: • Video: Traffic Cam, Static Cam 2 Vision (Video) • Social Media: Twitter Knowledge Graph • Text: Cambridge / WL City Data 3 Kafka Topics 1 Indexing Layer PostgreSQL Video ES Writer/Mapper Text 5 4 Front End User Profiling Microservice Users’ queries Triggers Heterogeneous Data Streams Active Learning Knowledge derived from queries Situation-aware Recommendation Situational Aware Indexed Data 41 41 Relevant patterns of data

Demo Video • In the demo video, we demonstrate as follows. • How twitter data is consumed and processed via Data Streaming Module • Extracting objects from Videos • Extracts the tweets that discusses about Object in Question • Tie features from different modality using the Indexing Layer • Build Index on the objects from videos and tweets • Functionality of the Front End with Graph Analytics • User Profiling extracts other objects that can be of users’ interest • Allows user to see those objects from all modalities 43

Demo Video • Simplified Query Select * from tweets, videos where tweets.objects_discussed == "car" videos.objects_detected == "car” • Demo Video URL – https://youtu.be/5TqWKzy5SqI 44

Feature Extraction Architecture ML, NLP Index Constructor Data type Processors NLP (Text) Knowledge Graph 2 Vision (Video) 3 Data Streaming 1 Indexing Layer PostgreSQL Kafka Topics ES Writer/Mapper Video Text 5 4 Front End Users’ queries Heterogeneous Data Streams Microservice Knowledge derived from queries Situational Aware Indexed Data 45 45 Relevant patterns of data

Extracting Features from Video with Deep Learning • Object detection and classification: best result achieved with deep learning architectures: – Faster RCNN – YOLO – SSD • Manual annotation and labeling – Time-consuming and expensive for large datasets – Outsourced human labor can be employed (MTurk) • We use pre-trained YOLO neural network to extract knowledge, detect and label objects in video • Retrain YOLO with Transfer Learning for detecting classes outside of pretrained ones 47

CNN-ROI based Architecture For Object Detection and Classification • YOLO detects 100+ classes • Our raw video dataset contains about 15 of the objects from these classes • YOLOv3 object detection algorithm 1. Regions of interests (ROI) proposals are generated 2. For each region, features are extracted and classified with Convolutional Neural Network 3. Apply non-maximum suppression: all candidate regions where probability of certain object detection is not max are dismissed 48

YOLO (You Only Look Once) Architecture 1. The image is split into an SxS grid of cells. 2. Each grid predicts B bounding boxes with C class probabilities • SxSxBx5 outputs in total 3. Conditional class probabilities are predicted Pr(Class(i)/Object): • SxSxC class probabilities • SxSx(B*5+C) output tensor • S=7, B=2, C=20 => (7,7,30) • Train a CNN to predict (7,7,30) tensor Image source: You Only Look Once: Unified, Real-Time Object Detection Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi https://arxiv.org/abs/1506.02640 49

YOLO (You Only Look Once) Architecture Objective : fast object recognition and detection Problem : CNN, R-CNN and modifications perform these tasks in multiple steps Solution : YOLO determines the object location and classifies it in one go • Optimal for streaming video • Input image is divided into SxS grid • Each grid cell predicts bounding boxes (B) and class probabilities (C) • Bounding box coordinates and class probabilities are encoded in an ouput tensor predicted by YOLO • Boxes with less than optimal Image source: You Only Look Once: Unified, Real-Time Object Detection Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi confidence scores are omitted after https://arxiv.org/abs/1506.02640 training 50

Detected Classes In the MIT Video Dataset CAR TRUCK PERSON BICYCLE TRAFFIC LIGHT PARKING STOP SIGN FIRE HYDRANT … AND MORE! METER 51

Preprocessing Tweets • Social media text has jargon, misspellings, special slangs, emojis 15:45 I luv my <3 iphone & you’re awsm apple, love you 🙐 3XXX. DisplayIsAwesome, sooo happppppy ฀ http://www.apple.com #apple @sjobs • Cleaning process – – HTML decoding – Expanding Contractions – Removing URL, Emoji, Reserved words, Smiley, User-mentions (or replace), hashtags • Preprocessing before tokenization – Remove punctuation, space, stop word 55

Additional tasks for Social Media Texts • Normalization of Noisy Text • Awsm ~ awesome, luv ~ love • Methodologies 1. Lexical normalization 2. Normalization with edit scripts and recurrent neural embeddings 3. Find balance between precision and recall 56

Topic Modeling with Tweets • Given a query keyword, we want to find similar tweets • We can do that by finding latent topics in tweets • Approach: Latent Semantic Analysis, or LSA • Raw counts do not work well as they do not account for the significance of each word in the document – ‘a’ has little significance in determining topic • Instead calculates the tf-idf score, w i,j – Takes the number of documents the word appears in into consideration • Document-term matrix is very sparse • So Dimensionality reduction is performed with SVD (Singular Value Decomposition) • From Document-term matrix, A, we retrieve – Term-topic matrix, V – Document-topic matrix, U • Document is represented with Term-topic matrix • Finally, Apply cosine similarity, sim(q, d) to evaluate:

Topic Modeling for Ontologies (Generative Models) • Even though LSA finds similar documents to user query, it has less efficient representation for topics. • Topics are necessary for ontologies while building our knowledge graph • LDA (Latent Dirichlet Allocation) – Generative Model – Uses Dirichlet priors for the document-topic and word-topic distributions – Results in better generalization for new documents – Allows online learning 62

Results: Similar Documents to Query Exact Key (93% TREE DOWN PERSON WITH GUN Similar) U 1 U 2 MASSACHUSETTS AVE SCOTT ST 64 64

Results: Similar Documents to Query Relevant Key TREE DOWN PERSON WITH GUN (70% Similar) U 1 U 2 MASSACHUSETTS AVE SCOTT ST 65 65

Multiple Data of Interest to Different Users Crime Disaster • Extract human-interpretable U 3 U 2 U 1 T T A T B c topics from a document corpus • Each topic characterized by Abduction Traficking Flood Earthquake words most strongly associated Streaming Theft Vandalism Rain Bush Fire Data with Others Others • Documents as mixtures of D 16 D 16 topics that spit out words with D 0 D 1 D 2 D 3 certain probabilities. Data at Rest D 31 • Uses variational Bayes for D 4 D 5 D 6 D 7 D 31 inference, no need to re-train D 8 D 9 D 10 D 11 D 43 D 12 D 13 D 14 D 15 D 43 69

Further Extension • Twitter data has metadata • Metadata bears a lot of information • Metadata can be used as context • Lda2Vec leverages a context vector to make topic predictions • We will adapt lda2vec • Context they used : sum of the word vector and the document vector https://multithreaded.stitchfix.com/blog/2016/05/27/lda2vec/ 70

Feature Extraction Architecture ML, NLP Index Constructor Data type Processors NLP (Text) Knowledge Graph 2 Vision (Video) 3 Data Streaming 1 Indexing Layer PostgreSQL Kafka Topics ES Writer/Mapper Video Text 5 4 Front End Users’ queries Heterogeneous Data Streams Microservice Knowledge derived from queries Situational Aware Indexed Data 71 71 Relevant patterns of data

Knowledge Graph ( Need to learn from ISI research) • Ontologies / Concepts are extracted from LDA • Extract Triplets <Subject, Relation, Object> to represent Events • Entities are represented by Nodes • Entities have Attributes (Labels) • Entities are connected by Relations (Edges) 72

Work In Progress with KG: Multi-modality  Multi-modal Information Retrieval  Poster represented In Northrop Grumman University Research Student Poster Competition 73

Feature Extraction Architecture ML, NLP Index Constructor Multimodal Streaming Data type Processors Data NLP (Text) Data Sources: • Video: Traffic Cam, Static Cam 2 Vision (Video) • Social Media: Twitter Knowledge Graph • Text: Cambridge / WL City Data 3 Kafka Topics 1 Indexing Layer PostgreSQL Video ES Writer/Mapper Text 5 4 Front End Microservice User Profiling Users’ queries Triggers Heterogeneous Data Streams Active Learning Knowledge derived from queries Situation-aware Recommendation Situational Aware Indexed Data 74 74 Relevant patterns of data

User Modeling: Intention-aware Recommendation Engine • Sends users streaming data that corresponds to their interests • Builds User Profiles using the history of user queries • Active Learning to narrow/expand intention model with more interaction • Expands user queries with word embedding models to fetch relevant data from the database Active Cars of specific make & model (purple Cadillac) • Analyze Learning to Interested in info. about crimes in a specific district • user User1 improve queries for SELECT * FROM video_data WHERE object = ‘car’ and intention user attribute=‘purple’ model with profiling time SELECT * FROM crash_data WHERE date_hit = TODAY User2 Expand result of Looks for pedestrians in the video data • queries with Interested in traffic, accidents, violations • word2vec 77

Feature Extraction Architecture ML, NLP Index Constructor Data type Processors NLP (Text) Knowledge Graph 2 Vision (Video) 3 Data Streaming 1 Indexing Layer PostgreSQL Kafka Topics ES Writer/Mapper Video Text 5 4 Front End Users’ queries Heterogeneous Data Streams Microservice Knowledge derived from queries Situational Aware Indexed Data 78 78 Relevant patterns of data

Data Streaming Module • Retrieve RESTFUL and Streaming Tweets • Populate Postgres with all data • Parse collected metadata to extract targeted information and store in Postgres • Replicable, fault tolerant, scalable and continuous • Build a Data Processing Pipeline with all features 79 79

Data Processing Pipeline Cambridge Public data (DB, CSV …) Parser Data Extraction Engine Engines Video Data Twitter Search API Twitter Topic Twitter Streaming API #Hashtag @User Profile Twitter 80 80

Retrieve Tweets : Implementation Choices • Search tweets by – Keyword / Hashtag (i.e, CambMA) – User Timeline (i.e, CambridgePolice) 81

Compatibility with other sources of data • Add new sources – JDBC – From file – Audio • Kafka Connect provides a framework (extra layer between source and Kafka) to develop connectors importing data from various sources and exporting it to multiple targets • Kafka Clients allow us to pass and retrieve messages directly to and from Kafka 82 82

Serving the Community • Collaboration with West Lafayette Police Department • Another Novel Use-Case • Extending SKOD Framework • Digs Deeper into Features, Knowledge Base and User Profiling

Extract targeted search results from heterogeneous data (i.e., video, police dispatch reports, Problem social media ) at rest and deliver Definition relevant information from incoming data streams based on context awareness.

Solution Framework Data Ingest System Insert Alert Postgres Video Feature Build Query from Identification Derived Features Incident Report Query System User Interface to Postgres Query Shows Current Result Display Query Results Postgres Trigger shows Future Results matching the current criteria 94

Police Dispatch Report

Incident Report to Features Incident Querying System Identified 31  Includes UI for entering police query for Features after fetching related information interviewing Sargent  Videos, Green  Similar Incident Reports, and Describing Suspect Attributes  Social Information  Functions as a data collection module 96

 Inquired features are input into Incident Report table in Postgres  From these features, system Incident builds Report to • Postgres Query: For fetching Search existing videos and reports matching the criteria Results • Postgres Trigger: Created for fetching incoming videos and incidents which will match the criteria

Feature Extraction from Incident Report Approach 1: Build Parsing Incident Regular expressions to Report for automatic filter out features and extraction of Features feature values. Assumptions - Beneficial for creating Report follows Postgres Triggers to Reports are highly identify similarity of grammar and is regular incoming incidents correctly structured to previous queries 99

Feature Extraction from Incident Report Approach 2: Build separate classifiers for each features separately and build an ensemble of classifiers • Latent Find related sentences that Semantic mentions selected features Analysis (LSI) Classifiers based on each • SVM , DTL features separately Improve upon • Embeddings basic BoW • BERT[2], Glove[1] features with • PoS, Feature Names 100

Approach 3: Formulate as Reading Comprehension Problem Feature Extractio Formulate Questions Formulate a fixed n from based on the structure for the selected features incident report Incident Report

Result: As Machine Comprehension

Approach 3 Challenges: Modifications to the model (To be worked on) - Binary answer - Span as an answer Is suspect wearing Jeans? - Does not recognize multiple feature mentions - Find each iterations of the answer 104

Jeans/Pants, Shoe, Color Hat, Shirt/Jacket, Segmentation Hair Color Feature DNN classifier trained for Male/Female Identification custom classes from Videos Walking/ Action Running/ Recognition In Pursuit

Search for Suspect in the Video • Matching the features extracted from text with features extracted from video Relate objects with similar features in Report Text and surveillance video 106

Surveillance Camera Video Datasets • Videos from HD surveillance cameras in main streets and high-traffic areas (WL) • Busy streets close to campus • Bar district with over 10 cameras down the State street • Over a month of archived video from dozens of cameras 107

Data Annotation • Yolo video processing tool — Yolo_mark ** – Automatically sample the video file into images – Manually label the images and output the label in text files – Compatible with Darknet framework • Used Yolo_mark to process video files at 1 frame / sec. • Manually identify images containing persons first • Depending on the “person”, label more refined attributes, e.g. male/female, jeans/pant, hat etc. – Extensive dataset needed for training (1000+ examples/class) • For compensation of different locations, angles, distances, time and weather conditions, 1/100 frames are chosen for annotation ** https://github.com/AlexeyAB/Yolo_mark 108

Yolo v3 with 53 layers [6] https://medium.com/@hirotoschwert/reproducing-training-performance-of-yolov3-in-pytorch-part1-620140ad71d3

Small Objects Detection • In surveillance video, objects tend to be of small size relative to the captured image • Short cut connections to skip layers are used for better detection of the small objects • The output of a shortcut layer is obtained by adding feature maps from the previous layer and 3 rd layer backwards • Shortcut connections strengthen feature propagation, reduce the number of feature maps to increase generalization 110

Color Segmentation • Goals: – Features: Jeans/ Jacket/ Shirt/ Hat – Values: Black Jeans / Blue Shirt – Avoid overhead DNN computation • Modified Yolo boxing codes to segment person into – head, upper half, bottom half, and foot • Sampling position is important to not include background colors

Detecting Clothes Color by Segmentation Analysis

Detecting Clothes Color by Segmentation Analysis

Result Frames

Result Frames

Color Segmentation • Bottom body is trickier due to different gestures and position • Run color analysis on different parts to extract color attributes – multiple shades of the same color, – night-time video colors are different • Rule Based Final Judgement of Color – Dark-colored or Light-colored – Multiple Shades of same color

Custom Training YOLO • Goals: Gender/ Race/ Wearables • Training on custom labelled dataset • Re-train with darknet** • Results are combined with color-segmented attributes * https://pjreddie.com/darknet/yolo/ 117

Video Data Processing Module • Raw video data is split in VideoID Extracted Location Timestamp Features/Color Coordinate 1-minute segments s s • Each segment is stored 1 ‘Male’, ‘Female’, 40.423994, - 11:35 AM, in RDBMS ‘Red Jacket’, 86.909224 15-Nov-2019 ‘Green Jacket’, • Each segment is ‘White Pants’ processed by custom- trained DNN and color 2 ‘Female’, ‘Red 40.423994, - 11:36 AM, segmentation module Jacket’, ‘White 86.909224 15-Nov-2019 Pants’ • Extracted features are 3 ‘Male’, ‘Female’, 40.423994, - 11:35 AM, stored in RDBMS with ‘Red Jacket’, 86.909224 15-Nov-2019 links to corresponding ‘Black Hat’ video segments 118

Video segments with the requested Query Results features are displayed: Dispatch Report Query searching for features: Suspect gender=female, jacket=true, jacket color=red, incident_date= '2019-11-15' incident_time= '20:00:00' Person with the searched attributes:

Query Result UI

Query Result UI

Deliverables and Demo GitHub Repository: Demo https://github.com/sko Incident Querying System - d-ng http://18.191.242.90/inde x.php Query Result UI - http://35.239.251.13:3000 /

http://35.239.251.13:3000/ Video samples extracted

REALM Incident Querying System For Policeman http://18.191.242.90/index.php

More SKOD Benefit Scenarios • Inform Drivers about – relevant obstacles and hazards: road closures, potholes, fallen trees and tree branches, ice, dumpster violations, downed road signs, not working traffic lights; – routes to avoid obstacles and hazards; – relevant POIs; – collision probability for a given date, time, weather conditions; recommend the speed. • Inform blind / differently abled people via a mobile app about: – relevant obstacles and hazards; – routes to avoid obstacles and hazards; – relevant POIs. 129

More SKOD Benefit Scenarios • Inform Law Enforcement about – suspicious activity (especially in crime-prevalent areas), illegal road constructions, downed road signs, blocked sidewalks, graffiti; – relevant obstacles and hazards; – routes to avoid obstacles and hazards; – collision probability for a given date, time, weather conditions; recommend the speed; – detected human faces in crime incidents and car accidents; – homeless people detected in certain areas. 130

References for WLPD 1. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global Vectors for Word Representation 2. Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina. 2018. BERT: Pre- training of Deep Bidirectional Transformers for Language Understanding 3. Seo. M et al (2017) Bidirectional Attention Flow for Machine Comprehension, in ICLR 2017 4. Paszke, A.; Gross, S.; Chintala, S.; Chanan, G.; Yang, E.; DeVito, Z.; Lin, Z.; Desmaison, A.; Antiga, L.; and Lerer, A. 2017. Automatic differentiation in pytorch. 5. AllenNLP: A Deep Semantic Natural Language Processing Platform. Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson Liu, Matthew Peters, Michael Schmitz, Luke Zettlemoyer. 2017 6. Yolov3 : An incremental improvement. J Redmon, A Farhadi 7. R-C3D: Region Convolutional 3D Network for Temporal Activity Detection - H. Xu et al, arXiv2017.

Selected Interesting References • ReXCam: Resource-Efficient, Cross-Camera Video Analytics at Scale, S. Jain, X. Zhang et al. https://arxiv.org/abs/1811.01268 • You’re being watched: there’s one cctv camera for every 32 people in uk. https://www.theguardian.com/uk/2011/mar/02/cctv-cameras-watching-surveillance. Accessed:2018-10-27. • Absolutely everywhere in beijing is now covered by police video surveillance. https://qz.com/518874/. Accessed: 2018-10-27. • Can 30,000 cameras help solve chicago’s crime problem? https://www.nytimes.com/2018/05/26/us/chicago-police-surveillance.html. Accessed: 2018-10-27. • https://github.com/PurdueCAM2Project (Prof. Yung Hsiang Lu at Purdue) • https://www.cam2project.net/ • Cross-dataset Training for Class Increasing Object Detection, Y. Yao, Y. Wang et al. https://arxiv.org/abs/2001.04621 • http://usc-isi-i2.github.io/knoblock/, • https://usc-isi-i2.github.io/kejriwal/ • PROTECTING AMERICA’S SCHOOLS A U.S. SECRET SERVICE ANALYSIS OF TARGETED SCHOOL VIOLENCE, 2019, • https://www.policyinsider.org/2019/10/protecting-americas-schools-a-us-secret-service-analysis-of-targeted-school-violence.html

133