Neuromorphic object detection and recognition

A car that can drive itself or a robot that can help with a wide variety of functions ranging from tedious housekeeping to dangerous military flight operations is the stuff of science fiction. Today's robots certainly make certain tasks easier but they are still incapable of functioning effectively for long periods in real-world scenarios (as opposed to controlled environments) without explicit human involvement [l, 2]. Robots that are intelligent need to be able to interact autonomously with objects in their surroundings [3]. This autonomous interaction involves three steps: (1) detecting the presence of the object [4], (2) recognizing the object—figuring out what type of interaction to have with it, and (3) tracking the trajectory of the object—determining how and when to react to it [4]. While these steps are computationally difficult, humans and many other living creatures are able to perform them easily. They are able to rapidly and effortlessly identify and categorize diverse objects in cluttered scenes under widely varying viewing conditions, such as changes in position, rotation and illumination [5]. Engineered systems are unable to match the level of proficiency and speed of biological visual systems. Through this thesis, we move towards the development of an autonomous, continuous-time visual system that emulates visual information processing in the primate visual cortex. This multi-stage system will utilize large-scale arrays of identical silicon neurons to build a biologically-plausible model of its biological counterpart. In particular, we have made advances in three areas. First, we have designed a neural array transceiver, the platform on which we intend to implement artificial cortex. This transceiver utilizes the Mihalas-Niebur neuron model, a generalized model of the leaky integrate-and-fire neuron with adaptive threshold, which is capable of producing most of the known spiking and bursting patterns of cortical neurons [6]. Ours is the first circuit implementation of the model and we have demonstrated many of the expected spiking patterns. Second, we implemented silicon facsimiles of cortical simple and complex cells according to the hierarchical model of object recognition in primate visual cortex proposed by Riesenhuber and Poggio [7]. We showed that the tuning and MAX computations at the V1-V2N4 stages of processing can be performed with integrate-and-fire neurons. Finally, we designed a neuromorphic cross-correlation engine, the first in neural hardware, for object detection. This is based on the work of Jonathan Tapson which suggests that correlation information can be obtained from the Interspike Interval Histogram (ISIH) of a spiking neuron [8]. On the cross-correlation engine, we were able to detect the auto and cross-correlation of signals as a mechanism for finding the content of information within the signal. In addition, we explored neural implementations of Kalman filtering as an object tracking mechanism and obtained encouraging preliminary results. In future, these functionalities—detection, recognition, and tracking—will be implemented on one or more neural arrays working together cohesively to produce a system with a potentially more intelligent artificial visual cortex.