Design Of Lightweight Object Classification System-on-Chip For Dynamic Vision Sensor

The frame-based vision processing systems have been widely used in various applications such as safety monitoring and assisted driving,and have achieved great success.However,frame-based vision sensors have the disadvantages of large output data redundancy,high output delay,and high energy consumption,and it is difficult to meet the high time efficiency and high energy efficiency requirements of high dynamic object recognition applications such as edge applications.Dynamic Vision Sensor(DVS)is a new type of bio-inspired vision sensor that uses Address-Event Representation(AER)as its data output and has the characteristics of data sparseness and low latency.The DVSbased visual processing system has the potential for higher time and energy efficiency.However,the current research on DVS-based visual processing systems is not mature enough,mainly using spiking neural network algorithms and general-purpose brain-like processor hardware platforms,which suffer from high system complexity and high hardware redundancy.Aiming at the application requirements of high dynamic object recognition,the thesis combines AER sensors and traditional machine learning technology to carry out the research of algorithm optimization methods for circuit realization,completes the optimization of the lightweight AER object classification algorithm,and completes the DVS-oriented lightweight object classification chip.The system has implemented on the Xilinx Zynq-7045 platform and achieved high throughput,low cost,and high energy efficiency.The main contributions of the thesis are as follows:(1)For the computational complexity of the DVS-oriented Spike Neural Network algorithm,the thesis adopts the lightweight AER object classification algorithm,and takes the logarithm of classconditional probability of the lightweight AER object classification algorithm,and uses the fixed-point operation and the look-up table method to optimize the lightweight AER object classification algorithm,which simplifies inference of the algorithm,makes the algorithm easier to implement by hardware;(2)In view of the short output time interval of AER events of DVS and the high real-time performance requirements of embedded devices,the parallel processing of AER events and multi-level pipeline technology,designed a lightweight AER object classification system-on-chip architecture with highspeed processing capabilities,which improves the system throughput and meets the realtime requirements of embedded applications;(3)The design and implementation of the on-chip online learning circuit of the system-on-chip has improved the flexibility of the system.After testing,the object classification accuracy of the system on the three data sets of MNIST-DVS,Poker-DVS,and Posture-DVS reached 77.9%,99.4%,and 99.3%respectively,which has certain application value.The system’s AER event throughput reaches 100 Meps,and the energy efficiency can reach 145 events/μJ.