S. Ghiasi, K. Nguyen, and M. Sarrafzadeh (USA)
Collaborative Processing, Accuracy-Latency Tradeoff, Profiling, Object Tracking
Various implementations of a tracking algorithm 0 5000 10000 15000 20000 25000 30000 35000 40000 45000 50000 4 6 8 10 12 14 16 18 20 Tracking accuracy (number of tracked features) Latency(ms) A B C D E Figure 1. Accuracy (or other design metrics) can be compromised for performance in constrained systems. Many parallel and collaborative signal processing systems utilize commercial off-the-shelf sensor nodes with constrained embedded processors. Applications running on such processors, e.g. object tracking, often demand real time performance and hence, another design metric such as accuracy has to be compromised to meet the performance constraint. Therefore, exact accuracy-latency characteristics of the application are required in order to implement it in a real time and sufficiently accurate fashion. This paper presents profiling techniques that are applicable to tracking applications including those implemented on a parallel system. The approach has been applied to a tracking application implemented on a collaborative system that has been built in our lab. Extensive profiling has been performed to study embedded vs. centralized and accuracy vs. latency tradeoffs. Experimental results verify the effectiveness of our profiling scheme, and support the fact that different computing schemes are appropriate for different accuracy and performance requirements. Experiments show that an appropriate choice of algorithms and computing schemes of our system, leads to 12 times speedup in feature tracking latency compared to its original version with a reasonable reduction in tracking accuracy. On the other hand, many applications that utilize constrained sensor nodes demand real time performance. Examples include multimedia and tracking applications that use cameras with low power embedded processors. Such applications cannot tolerate slow processing and long latencies since they have to process the real time stream of incoming data. Therefore, designers have to compromise one of the design metrics, such as accuracy to obtain the desired performance for this class of applications.
Important Links:
Go Back
