Section: New Results

Highlights of the Year

CovTrack: Agile multi-target multi-threaded realtime tracker

We have developed and highly optimized a multi-target tracking system based on covariance tracking algorithm. The complexity of the algorithm – connected to the number of features – can be tuned to fit the processor computation power (with/without SIMD). Moreover the features can be also selected from a large set of features to adapt the algorithm to the scene and the nature of tracking (indoor/outdoor, pedestrian/car, ﾉ). Some software and algorithmic transforms have been also applied to accelerate the code for scalar/SIMD processors. [20]

The Light Speed Labeling (LSL) algorithm

is still the world fastest connected component labeling (CCL) algorithm. We have proposed a new benchmark that performs fair comparisons for such a data-dependent algorithm (that involves Union-Find algorithm optimization combined with memory and control flow optimization). We show that thanks to its run-based approach and its line-relative labeling, LSL is intrinsically more efficient that all State-of-the-Art pixel-based algorithms, whatever the memory management.[23]