First results of a detector embedded real-time tracking system with artificial retina

Abstract

We present the first results of a tracking system prototype using the artificial retina algorithm for fast track finding. The system is based on extensive parallelization and interconnectivity, and allows real-time tracking with offline-like quality with a latency < 1 μs. The artificial retina algorithm has been implemented on a novel custom data acquisition board, based on commercial FPGAs, that we have designed and constructed. The retina architecture is organized in three main blocks: a switch for the parallel distribution of the hits, a pool of engines for the digital processing of the hits, and a block for track parameter calculations. In the switch module memory buffers store the hits information according to a hold logic that gets activated when downstream modules are busy. A real-time tracking prototype has been built and tested successfully on beam at the CERN SPS with input track rate of about 300 kHz. It consists of a telescope with 8 planes of single-sided strip detectors readout using custom ASICs and providing hit position to the tracking processor. The tracking performance of the system are compatible with offline results and the system response is in agreement with simulations

Event-Based Sensing and Signal Processing in the Visual, Auditory, and Olfactory Domain: A Review

“…An extensive parallel tracking system is designed in Neri et al (2015), that allows real-time tracking withe a latency of <1 µs. The retina architecture is organized in three main blocks.…”

“…After 2006, there is steady increase in the research on the subject, indicating that the potential of the technology has been discovered. (Etienne-Cummings and Van der Spiegel, 1996;Delbruck and Liu, 2012;Posch et al, 2014;Gallego et al, 2020), Period of 2002(Delbruck, 2016, neuromorphic chips , VLSI neuromorphic circuits (Indiveri, 2008), spiking neural networks (Brette et al, 2007) Benchmarks Guidelines for benchmark creation (Tan et al, 2015), dataset (Gibson et al, 2014a;Li et al, 2017), object recognition (Serrano-Gotarredona and Linares-Barranco, 2015), action recognition and tracking (Hu et al, 2016), 3D perception (Zhu et al, 2018), driving applications (Binas et al, 2017) Applications Tracking Object tracking (Gómez-Rodríguez et al, 2011;Saner et al, 2014;Delbruck et al, 2015;Zong et al, 2018), multiple object (Gómez-Rodríguez et al, 2010;Linares-Barranco et al, 2015), camera movement (Kim et al, 2008;Reinbacher et al, 2017), feature tracking (Lagorce et al, 2015b;Ni et al, 2015;Alzugaray and Chli, 2018a), stereo tracking (Schraml et al, 2010b;Müller and Conradt, 2012), camera pose (Gallego et al, 2015(Gallego et al, , 2016(Gallego et al, , 2018aMueggler et al, 2015c), micro-particle tracking (Drazen et al, 2011;Borer et al, 2017), subatomic particle tracking (Neri et al, 2015(Neri et al, , 2017, car tracking (Litzenberger et...…”

Real-time tracking system using precise space and time information of the hit

“…Using 512 cellular units for the grid of track parameters with a granularity of 3.3 mm the response of the algorithm was simulated at a clock frequency of 200 MHz. The choice of 512 engines is motivated by an existing implementation of the artificial retina algorithm with no time information [12] and found useful for comparing the utilization of resources in the two different cases. The fast track finding architecture based on the artificial retina algorithm is composed by three different modules: the switch module that delivers the hits from the detector layers to appropriate For the implementation of the 4D artificial retina algorithm, minor changes have been made to the switch and the track fitter logic modules with respect to the artificial retina algorithm implementation using spatial information only.…”

“…The multiplications between space and time Gaussian responses are implemented using DSP (digital signal processing) blocks. In the track fitter the x − and x + track parameters are evaluated, as for the case of the artificial retina algorithm, by interpolation of the weight values of the cells adjacent to the local maximum [12]. These modifications to the architecture of the system require a modest increase of FPGA resources of about 10%.…”

“…Using 512 cellular units for the grid of track parameters with a granularity of 3.3 mm the response of the algorithm was simulated at a clock frequency of 200 MHz. The choice of 512 engines is motivated by an existing implementation of the artificial retina algorithm with no time information [12] and found useful for comparing the utilization of resources in the two different cases.…”