Real-Time Cyber Physical Systems Application on MobilityFirst Winlab Summer Internship 2015 Karthikeyan Ganesan, Wuyang Zhang, Zihong Zheng Shantanu Ghosh, Avi Cooper
WINLAB SUMMER 2015 TEAM M MEMBER BERS Karthi hikey eyan an Ganes esan an Wuyang ang Zhang ng Zihong ong Zheng ng Avi Cooper er Shant antanu anu Ghosh
WINLAB SUMMER 2015 PRE RELIM IMINARY INARY GOAL AL OF OUR UR PRO ROJECT ECT CPS Application based on MF Serv rver r side: e: Client ent side: de: Mobil ilityFi ityFirst rst Imple lement ment server ver Run an insta tance nce of Virtua tual Ne Networ work applicatio cation for object ct camera ra syste stem; m; recogn gniti ition on; Transmi nsmits ts video o in Return rn the result lt standard rd forma mat; t; Simple le graphical hical interf rface ce to display ay results lts
WINLAB SUMMER 2015 CU CURR RRENT NT FRA RAME Image Recognition: Done essential time analysis for different phases in Image Matching. Have some progress on Speed optimization. Also working on strategies to improve Accuracy. Cloud Computing: Set up the Hadoop Cluster, constructed by Master and Slaves that could run Map & Reduce jobs. Now working on Hadoop Image Processing Interface(HIPI). Application: Tried to debug some MF Android applications done before(mfstack, mfping, etc.) Learning how to set up the WIFI access point on Router to enable the MF connection.
WINLAB SUMMER 2015 IMAG AGE PRO ROCES CESSO SOR • Server Startup • Load Descriptors into memory • Image Matching • Isolate Descriptors from test image • Find Matching Descriptors from Database using the Flann Class • Algorithm to determine Object from matched descriptors
WINLAB SUMMER 2015 IMAG AGE PRO ROCES CESSO SOR R OPTIM TIMIZ IZATI ATION: N: SPEED • Server Startup • Load Descriptors into memory • Image Matching • Isolate Descriptors from test image • Find Matching Descriptors from Database • Build a KD tree • Perform a knn search • Algorithm to determine Object from matched descriptors
WINLAB SUMMER 2015 IMAG AGE PRO ROCES CESSO SOR R OPTIM TIMIZ IZATI ATION: N: SPEED Old New 4500 40 4000 35 3500 30 ms) 3000 ms) h Time (ms h Time (ms 25 2500 20 2000 Search Search 15 1500 10 1000 500 5 0 0 50 100 250 500 1000 50 100 250 500 1000 Image Databas ase e Size Image Databas ase e Size
WINLAB SUMMER 2015 IMAG AGE PRO ROCES CESSO SOR R OPTIM TIMIZ IZATI ATION: N: SPEED • Descriptor Isolation • Largest portion of computing time. • Fixes : • Try utilizing the GPU to perform SURF • Try a method that better utilizes multiple CPU cores • Try different feature detector other than SURF
WINLAB SUMMER 2015 IMAG AGE PRO ROCES CESSO SOR R OPTIM TIMIZ IZATI ATION: N: ACC CCUR URACY CY • True Positives • Detects objects in test image that exist in Database with close to 100% accuracy • False Positives • Often Falsely recognize objects that don’t exist in the image (Close to 50%) • Fix • Improve the Algorithm used to determine object from Matched Descriptors with additional checks to confirm the actual existence of the object.
WINLAB SUMMER 2015 ANDROID PHONE MFPing achieves the MFStack is used to basic ping function for install the MF stack on MF such as the ping for phone. TCP/IP. Also a launcher to start We plan use it to test the and stop the mf stack MF connection after we service. set up the access point on router. Devices under MF network is actually communicating through the MF stack.
WINLAB SUMMER 2015 PLAN TO SET UP THE CONNECTION Glass captures video ORBIT outdoor nodes HADOOP Slaves nodes Bluetooth MF MF Android phone handles Master node as Server Access Point on router networking manages/allocate image recognition jobs
WINLAB SUMMER 2015 HIPI FRAMEWORK Hadoop Image Processing Interface Background: Optimized image processing algorithm, current system process 1,000 images within 113 ms. Impediment: Load the large size of descriptors in the database into a local memory before matching images. 1,000 images : 1 GB, linear accumulation -> 10,000 images : 10 GB HIPI: Image processing interface under the framework of HADOOP. Distribute the database into several machines so as to support over 10,000 images.
WINLAB SUMMER 2015 Ne Next t Week k Plan Set et up the he conn nnection ection bet etwee een n And ndroid oid Pho hone ne an and orbi bit t nodes through gh an access ss point. t. Conti tinue develop p the client progra gram based d on google gle glass ss and Androi oid d phone. Use e HI HIPI I to em embed edde ded d Image age Proc ocess essin ing Prog ogra ram m an and Ha Hadoo oop p Frame mework. work.
Questions?
Recommend
More recommend