2019/04/24 by Bharath Ramesh, Ramesh, Bharath, Andrés Ussa +7
Engineering · Materials Science · #Advanced Memory and Neural Computing #Ferroelectric and Negative Capacitance Devices #Electronic and Structural Properties of Oxides
paper · pdf · doi:10.48550/arxiv.1904.12665
We present the first purely event-based, energy-efficient approach for object\ndetection and categorization using an event camera. Compared to traditional\nframe-based cameras, choosing event cameras results in high temporal resolution\n(order of microseconds), low power consumption (few hundred mW) and wide\ndynamic range (120 dB) as attractive properties. However, event-based object\nrecognition systems are far behind their frame-based counterparts in terms of\naccuracy. To this end, this paper presents an event-based feature extraction\nmethod devised by accumulating local activity across the image frame and then\napplying principal component analysis (PCA) to the normalized neighborhood\nregion. Subsequently, we propose a backtracking-free k-d tree mechanism for\nefficient feature matching by taking advantage of the low-dimensionality of the\nfeature representation. Additionally, the proposed k-d tree mechanism allows\nfor feature selection to obtain a lower-dimensional dictionary representation\nwhen hardware resources are limited to implement dimensionality reduction.\nConsequently, the proposed system can be realized on a field-programmable gate\narray (FPGA) device leading to high performance over resource ratio. The\nproposed system is tested on real-world event-based datasets for object\ncategorization, showing superior classification performance and relevance to\nstate-of-the-art algorithms. Additionally, we verified the object detection\nmethod and real-time FPGA performance in lab settings under non-controlled\nillumination conditions with limited training data and ground truth\nannotations.\n