Lightweight tracking-by-detection system for multiple pedestrian targets

摘要

In this work we present a tracking-by-detection system that works under two hard computational constraints: no parallelization and real-time performance. To this aim, we develop a lightweight method for object detection based on background subtraction, window masking and image density projections. The association of detections and trackers is carried out with the Hungarian algorithm on a loss matrix that is constructed according to position, size and appearance. In addition we include heuristics for detecting forming groups as well as incoming and outgoing pedestrians. Finally tracking is performed with particle filters. We test our system on different well-known benchmark datasets. Experimental results reveal that the proposed method is efficient and effective. Specifically, it obtains a processing rate of up to 22 frames per second on average when tracking up to 9 people.