Discovering frequent and rare spatio-temporal patterns in large amounts of streaming visual data is of great practical interest since it allows for automated applications of activity and surveillance analysis. In this paper we present a computationally efficient and memory preserving clustering scheme which uses streaming input from a stationary-mounted neuromorphic camera and performs density-based clustering in a high-dimensional feature space. The clustering scheme can treat arbitrarily shaped complex distributions and employs an intuitive density-based criterion to assign previously unseen samples to categories of frequently observed and rare. The spatio-temporal structure of neuromorphic video is encoded into sparse binary features, which allow for fast Hamming distance based neighborhood analysis in the feature space. Moreover, data sparsity brings advantages with respect to memory-efficient transmission and storage of the learned statistical model when used within a camera network. We present rare event detection results in a multiple-day neuromorphic data sequence and discuss strengths, failure modes and possible extensions of the proposed method.