Video Based Illegal Dumping Detection Alan Chen Marcus

Video Based Illegal Dumping Detection Alan Chen, Marcus Garcia, Nathan Wong, Hossein Shahdoost, Prem Bharath Soundararajan, Youngsoo Kim Department of Electrical Engineering San Jose State University 1

Outline • Project Objectives • Literature Survey • Project Status • Conclusion and Future Work 2

Project Objectives • Improve the accuracy and speed of automated detection of illegal dumping video • Accelerate its detection by 10 ~ 100 X by parallel execution • De-noising of videos using DWT (Discrete Wavelet Transform) thresholding techniques • Use of HOG (Histogram Oriented of Gradients) for increase accuracy of detection 3

Literature Review - Array of Things (Ao. T) • An urban sensing project with a network of interactive, modular sensor boxes around Chicago. • Goals: Monitor Chicago’s environment, infrastructure, and activity. • Reports real time and location based data about the city. • Applications: Health warning, population density/traffic, etc… Collaborative Project with the city of Chicago, University of Chicago, and Argonne National Laboratory • Prototyping, 50 nodes as of Summer 2016 http: //www. netlib. org/utk/people/Jack. Dongarra/CCDSC-2016/slides/talk 07 -beckman. pdf 4

Literature Review - Georgia Tech, Smart City Project • Pedestrian and car detection • Precedent-Aware Classification • Reduce unnecessary calculations through calibrated detectors • Aspect ratio, size, pixel density, frame location • Create a common pathway for increased detection • Test Results • Xilinx Zynq 7020 So. C • Dual ARM Cortex-A 9 and Integrated FPGA • 16% increase in accuracy over traditional object detection methods Common Pathway 5

Open Source Object Detection (Base • Platform) Obj. Left is a C++ program developed by the National Taiwan University to detect abandoned bags and identify owners. It uses Open. CV and Open. MP running on a Linux system. • Obj. Left uses background subtraction to identify background and foreground objects using 2 -bit pixel-based FSM. Dual-rate foreground integration is used to classify objects as background or static foreground objects. 6

Open Source Object Detection Obj. Left constructs background models using slow and fast learning rates. Comparison between the two background images indicate abandoned objects. 7

Project - Detection • Our Approach • ALDEC Ty. SOM-2 Embedded Prototyping Board • Xilinx Zynq 7000 • Dual Core ARM Cortex-A 9 • Camera(s) connected to a ALDEC board capturing frames. • Video clips can be input to ALDEC board for processing • Processing frames using OBJLEFT code to detect abandoned objects 8

Stand Alone Solution • Connection • Without Wi-Fi/Radio • Remote Hotspot • Bluetooth Device communication • Materials • Camera(s) • Box containing FPGA controller (ALDEC)/ Bluetooth device • Features • • Increase Accuracy Detection via HOG De-noising via Thresholding Garbage/Debris Detection Battery Powered (1 Month duration) 9

Multicores Acceleration [Youngsoo Kim, Marcus Garcia, Alan Chen, Nathan Wong, Shrikant Jadhav, and Clay S. Gloster, “Smart City Service Acceleration on FPGAs”, IEEE International Workshop on Big. Data Security and Services 2017] Obj. Left Average FPS per Open. MP Thread 30 25 Average FPS 20 15 10 5 0 1 2 CPU Threads 4 8 10

Discrete Wavelet Transform Processor for De-noising [Alan Chen, Marcus Garcia, Youngsoo Kim, “ 300 MHz Smart Camera DWT Processor” 11 TH INTERNATIONAL CONFERENCE ON DISTRIBUTED SMART CAMERAS, 2017] • DWT is a derivative of DCT. • De-contruct, De-noise, Re-construct • DWT is superior to DFT when considering time-frequency data analysis 11

Improving Accuracy for Detection [Youngsoo Kim, Hossein Shahdoost, Shrikant Jadhav and Clay S. Gloster, “Improving the Accuracy of Arctan for Face Detection”, The 25 th IEEE International Symposium on Field-Programmable Custom Computing Machines, 2017] • 12

Conclusion and Future Work • Accelerated up to 10 x from ~2. 5 fps to 25 fps • Hardware acceleration of up to 100 x • DWT for de-noising • Compression of video sequences • Advanced feature extracting for the City 13

Publications • Alan Chen, Marcus Garcia, Youngsoo Kim, “ 300 MHz Smart Camera DWT Processor”, 11 TH INTERNATIONAL CONFERENCE ON DISTRIBUTED SMART CAMERAS, 2017 (SUBMITTED) • Youngsoo Kim, Marcus Garcia, Alan Chen, Nathan Wong, Shrikant Jadhav, and Clay S. Gloster, “Smart City Service Acceleration on FPGAs”, IEEE International Workshop on Big. Data Security and Services 2017. • Youngsoo Kim, Hossein Shahdoost, Shrikant Jadhav and Clay S. Gloster, “Improving the Accuracy of Arctan for Face Detection”, The 25 th IEEE International Symposium on Field-Programmable Custom Computing Machines, 2017. 14