Boosted kernelized correlation filters for event-based face detection

Recently, deep learning has revolutionized the computer vision field and has resulted in steep advances in the performance of vision systems for human detection and classification on large datasets. Nevertheless, these systems rely on static cameras that do not yield practical results, especially for prolonged monitoring periods and when multiple object activities occur simultaneously. We propose that event cameras naturally solve these issues at the hardware level via asynchronous, pixel-level brightness sensing at microsecond time-scale. In particular, event cameras do not output data during no-activity periods and thus data rate is drastically lowered without any additional processing. Secondly, event cameras produce disjoint spatial outputs for multiple objects without requiring segmentation or explicit background modeling. Leveraging these attractive properties, this paper presents an event-based feature learning method using kernelized correlation filters (KCF) within a boosting framework. A key contribution is the reformulation of KCFs to learn the face representation instead of relying on handcrafted feature descriptors as done in previous works. We report a high detection performance on data collected using an event camera and showcase its potential for surveillance applications. For fostering further research, we release the face dataset used in our work to the wider community¹.