Hongrak Son

Optimal path finding with space- and time-variant metric weights via multi-layer CNN

Analogic CNN-based optimal path-finding algorithm is proposed to solve the problem with space- and time-variant metric weights. The algorithm is based on the analog version of modified dynamic programming which is associated with non-linear templates and multi-layer CNN employing the distance computing (DC), the intermediate (I), and the path-finding (PF) layers. The cell outputs of I layer are jointly utilized among the cells on the DC layer and the PF layers, which allows the network structure to be compact. The arbitrary levels of metric weights can be provided externally and the real-time processing of the optimal path finding is achieved on the space with the time-variant metric weight. Parallel-processing capability for the multiple optimal path finding is the additional property of the proposed algorithm. The proposed multi-layer CNN structure and its non-linear templates are introduced. The proper operation of the proposed structure is verified through theoretical analysis and simulations. Copyright

Analog addition/subtraction on the CNN-UM chip with short-time superimposition of input signals

The cellular-neural-network universal machine (CNN-UM) technique which performs analog addition/subtraction between image frames has been developed. The equivalent circuit of the uncoupled CNN without self feedback is reduced to a simple RC circuit. If two inputs are presented to the circuit one after another during a very short time period, the voltages that are proportional to their input signals are superimposed on the state capacitor. The output of such superimposition is a reduced version of the addition/subtraction between the two signals. Simple amplification of the output can recover the actual output. The characteristics of analog addition/subtraction with the proposed algorithm are shown via on-chip experiment. Application of the proposed algorithm to moving target detection is also presented.

Very high speed Viterbi decoder with circularly connected analog CNN cell array

A very high speed Viterbi decoder with a circularly connected 2-dimensional analog CNN cell array is disclosed. The circuit is circularly connected, forming a cylindrical shape so that the cells of the last stage are connected to those of the first stage. Such configuration enables the conceptually infinitive length of the trellis diagram to be implemented with a limited size of circuits. Use of the analog circuits does not require the necessity of the A/D converters. With the parallel analog processing structure, its decoding speed becomes very high. Features of the proposed Viterbi detector are compared with those of the conventional digital Viterbi decoder.

A depth measurement system with the active vision of the striped lighting and rotating mirror

A depth measurement system that consists of a single camera, a laser light source and a rotating mirror is investigated. The camera and the light source are fixed, facing the rotating mirror. The laser light is reflected by the mirror and projected to the scene objects whose locations are to be determined. The camera detects the laser light location on object surfaces through the same mirror. The scan over the area to be measured is done by mirror rotation. Advantages are 1) the image of the light stripe remains sharp while that of the background becomes blurred because of the mirror rotation and 2) the only rotating part of this system is the mirror but the mirror angle is not involved in depth computation. This minimizes the imprecision caused by a possible inaccurate angle measurement. The detail arrangement and experimental results are reported.

High speed road boundary detection on the images for autonomous vehicle with the multi-layer CNN

A multi-layer CNN-based algorithm to find the most likely road boundaries on camera images is proposed for the possible application to autonomous vehicle driving. In the previous study, the Dynamic Programming (DP) is shown to be implemented with the multi-layer CNN. If the road-edge images are treated as the space variant distance weights, the optimal path finding algorithm of CNN-based DP can detect the optimal road boundary. Partly disconnected boundary line segments of roads could be linked by way of the most likely road boundary line segments. Fast processing speed is another advantage of the proposed CNN-based structure if it is implemented with hardware circuits. Simulation results about various different road images are included.

Target Tracking via Region-Based Confidence Computation with the CNN-UM

A target tracking algorithm with the region-based confidence computation on the CNN-UM is proposed. The CNN-UM is an analog parallel computational system which handles regions easily with its region creating capability, parallel processing in the region and regional constraining capability. If the probability for each feature is created in each region, the total confidence of a target can be computed with a fusion algorithm employing products of weighted sums of feature probabilities. The cell-wise target decision in the region can be performed depending on the confidence value at each cell. By virtue of the analog parallel computational structure of the CNN-UM, the computation speed is very fast. On chip experimental results are included in this paper.

Dependant distance potential source algorithm for optimal path finding with the analogic CNN

The Dependant Distance Potential Source algorithm is proposed for the analog CNN-based optimal pathfinding in long distance path problems. Analog circuit-based dynamic programming is a promising solution for real-time optimal path finding. For easiness of its hardware implementation, the Distance Potential (DP) concept can be utilized, where the optimal path is determined on the DP field. Due to the limited dynamic range of the analog circuits and so DP field coverage, the DP concept alone does not give the complete solution for to the long distance path problem. The Dependant Distance Potential (DDP) algorithm is proposed to resolve this problem through extension of the DP field coverage by setting the enhanced DDP sources.

Robust optical flow detection based on the distance transform with the CNN nonlinear circuits

A robust optical flow computation algorithm utilizing the trajectories of feature points has been developed. For some applications of optical flows, correct optical flows (though they are not so many) are more useful than unreliable ones at every pixel point. The proposed algorithm is for detecting the optical flows only at the feature points. The optical flow vectors are extracted from the trajectory segments of feature points on which distance information is developed through a distance transform. A multi-layer cellular neural network (CNN) structure and nonlinear templates for the proposed algorithm are suggested and examined. Simulation results show that the proposed algorithm is robust against noise, even without any preprocessing.