The Design Of Multi-orientation Moving Object Recognition System For Bio-inspired Vision Sensors

The address-event representation image sensor is one of the bio-inspired vision sensors.Compared with traditional image sensors based on frame,it has the traits of low power consumption,low redundancy,low delay and real-time.The AER image sensor captures the moving objects in a high speed,which is suitable for object recognition and tracking in machine vision system.The majority of these event-based categorization systems only can recognize the objects moving in a specified orientation,which limits their application for classification.Increasing the training samples moving in multi-orientation is not benefit to perform,because it will add the workload of data acquisition and raise the additional training time dramatically.In order to recognize the objects moving in different orientations,we develop a new algorithm only using the single orientation objects as training samples.In addition,the inter-module parallel processing method is convenient for hardware implementation.It provides ideas for research on multi-orientation machine recognition system chip.Firstly,this paper briefly analyzes the concept,principle and application of the AER image sensors in different types including DVS,ATIS and DAVIS image sensor,and a conversion method between gray-level image and the events is proposed.In the preprocessing phase,a feature extraction algorithm based on the AER image sensor is designed.It contains convolution,forgetting mechanism and the MAX operation to extract multi-scale and multi-orientation features.In the meantime,an event-based recognition algorithm is proposed.The spiking neural network is selected to identify the responses from multiple feature extraction.Finally,an event-based tracking algorithm is combined with the above algorithms to design a multi-orientation moving object recognition algorithm.The feature extraction module and the orientation detection module work in parallel to realize the classification of moving objects in random directions.The experimental results show that the multi-orientation moving object recognition system based on the AER image sensor can realize the function of classification.The MNIST handwritten digit dataset,the poker dataset and the vehicle dataset are used as the experimental samples.The results are as follows: 1.The accuracy of orientation detection is more than 90%.The horizontal and vertical direction detection accuracy is above 99%.The error of vehicle orientation detection is less than 3o.2.The recognition accuracy moving in a single direction is superior to the existing algorithms.The recognition precision achieves 93.78% in the MNIST handwritten dataset experiment.3.Multi-orientation moving object recognition function can be realized.In the card experiment,the correct rate is 76.39% to recognize card moving in random directions.To sum up,the proposed multi-orientation moving object recognition system can efficiently classify the objects in any direction,and it is suitable for high-speed AER vision system applications.