Research On Key Techniques Of Multi-object Recognition System For Bio-inspired Vision Sensors

The Address Event Representation(AER)image sensor is one of the bio-inspired vision sensors with low data redundancy and suitable for high-speed target tracking and recognition.Because of the special ouput of AER sensors,which is a stream of events,it is necessary to develop a new recognition system based on events.At presents,most AER recognition algorithms presented focus on the accuracy on the ideal database,but the nonideal factors in practical applications are also of great research value.In this paper,a multi-object recognition system for bio-inspired vision sensors is studied,and the system optimization is designed for practical problems.This paper first analyzes the working principle of the AER image sensor,then proposes a multi-target tracking algorithm,which can evaluate the state of the target precisely.Based on the intermediate data of the tracking algorithm,an activity peak detector is designed to get the opening time of the classifier.Then,a feature extraction algorithm for AER images is designed,which adopts Gabor convolution kernels and maximum pool operations to achieve and compress the features in different directions and scales.At the same time,this paper proposes a spiking neural network based on the Tempotron learning mechanism and completes the preliminary development of the multi-object recognition system based on MATLAB GUI software design platform.Finally,in view of the the nonideal factors such as noise and multiscale input,based on the analysis of the existing methods,this paper proposes a double weighted centroid algorithm and a feature scaling algorithm to realize the optimization of the multi-object recognition system.The experimental results are as follows: 1.The multi-object tracking algorithm can detect and locate differecnt AER objects and the running time of the detector is deduced to 52.08%.2.The double weighted centroid algorithm can greatly reduce the interference of the noise,and the recognition precision achieves 87.64% on the MNIST handwritten dataset with 2.5% salt and pepper noise density,which is superior to the existing algorithms.3.The feature scaling algorithm can solve the multi-scale problem,and the recognition precision achieves 83.29% when the scales of AER images vary in a wide range.To sum up,the proposed multi-object recognition system has the performance of solving nonideal factors such as noise,multi-target and multiscale input,and is suitable for the practical application of high-speed AER vision system.