1995 TI DSP Solutions Challenge Regional Winners

Territory III (Asia/Japan Division)

• Western Japan
• Eastern Japan
• Asia

Western Japan

Entry Topic: 		"Realtime Person Tracking System"
University: 		Osaka University
Team Members: 		Ryuzo Okada, Graduate
			Shinya Yamamoto, Graduate
			Yasushi Mae, PhD Candidate
Advising Professor: 	Dr. Yoshiaki Shirai, Dept. of Mechanical 
			   Engineering

Abstract

We developed a real-time system for tracking a moving person by using a high speed image processor and an active camera head. The image processor is composed of five DSP boards, each of which include two DSPs. The two DSPs on the board can communicate with each other through the communication ports. The DSPs perform various image processing in parallel. The first three DSP boards calculate optical flows (a velocity field of an image based on the gradient method.) Next, the region of a person on the image is extracted from the optical flows using knowledge about the shape of a person. Finally, the camera head is controlled to keep the person in the field of view of the camera. The result of an experiment is shown for tracking a person in a cluttered background.

Eastern Japan

Entry Topic: 		"Development and Applications of 
			   Multi-DSP System"
University: 		Musashi Institute of Technology, 
Team Members: 		Hideaki Matsumoto, PhD Candidate
			Toru Nakayama, Graduate
			Dai Haruki, Graduate
			Kenichi Yoshii, Graduate
			Masashi Yoshida, Graduate
Advising Professor: 	Dr. Mototaka Sone, Professor, Dept. of 
			   Electrical Engineering

Abstract

We have produced a multi-DSP system, designed by using simulation, a C compiler, a task scheduler, and software applications for efficient parallel processing, in order to perform super high-speed computation. We have been researching the electric power system, finite element analysis and various control systems. Since it takes much time to compute them, the super high-speed computation is very much required. However, DSP multiprocessor and software applications which can realize high-speed computation were not marketed. Therefore, we had to develop both the Multi-DSP System and software applications. At the design stage of this system, we decided the most suitable number of processors and network construction by simulation. From this result, we used both the global bus network and the binary 5 cube network with 32 processors. In order to get high performance, we selected the TMS32OC40. Moreover, to get rid of the global bus conflict in shared memory, a TMS32OC31 as a control processor was used. In this way, we succeeded to produce a multi-DSP system using 32 TMS32OC40s with a total of 1.28 GFLOPS performance.

In software development, it is difficult to program by DSP assembly. Therefore, in order to make full use of DSP's unique architecture such as plural arithmetic units, plural memory block and hardware loop, we developed new C-like language and a compiler. In this case of using multiple processors, we needed a parallel algorithm. To solve the problem we developed a task scheduler using the Critical Path method too.

Using this system stated above, we accelerated two important problems; power system analysis and finite element analysis. In order to speed up a transient stability analysis of power system, we divided a job of the analysis into tasks, and the tasks were processed in parallel by the Multiple Instruction Multiple Data model. As a result, it was able to analyze more than 300 generators in real-time. In addition, since it is possible in an instant to compute, the use of this fast analysis method make possible predictive digital control.

Asia

Entry Topic: 		"High Performance Motion Picture 
			   Restoration System"
University: 		Nanyang Technological University, Singapore 
Team Members: 		Showbhik Kalra, Undergraduate
			Dilip Krishnan, Undergraduate
Advising Professor: 	Dr. Chong Man Nang, Lecturer, School of 
			   Applied Science

Abstract

This project presents a computationally efficient, parallel motion picture restoration system to restore old motion picture archives. A novel Gaussian Weighted, Bi-directional Auto-Regressive 3-Dimensional (GWBAR-3D) algorithm is formulated to alleviate the presence of noise in the old archives. Common forms of degradation found in such archives are "dirt and sparkle" and scratches. The distortion is caused either by the accumulation of dirt or by the film material being abraded.

While most of the existing image restoration algorithms will blur edges of moving objects in the vicinity of occluded and uncovered image regions, the proposed GWBAR-3D algorithm is able to suppress mixed noise processes and recover lost signals in both the covered and uncovered regions in an image sequence. The proposed algorithm is tested on the artificially corrupted image sequences and real distorted video (full PAL image size).

In our original design of the parallel motion picture restoration system, we reduce the number of computations in the GWBAR-3D algorithm: the restoration process is interpreted as an interpolator instead of a LMS process. The computational efficiency of the GWBAR-3D algorithm is significantly improved where the improvement in the sequential processing speed alone is in excess of 100 times. The computationally efficient GWBAR-3D algorithm is parallel implemented on an array of TMS320C40 processors connected in a tree configuration. Two different parallel algorithms are implemented in which a close to linear speed-up is achieved by means of an improved algorithm. The scalability of the proposed 15-node motion picture restoration system is demonstrated.