Hardware Implementation And Performance Analysis Of Convolutional Neural Network For Video Pedestrian Detection

In recent years,convolutional neural networks(CNN)have been widely used in pedestrian detection.In order to implement video pedestrian detection algorithms based on CNN to practical applications,some problems of algorithm design,hardware selection and algorithm implementation need to be solved in engineering.The implementation and performance analysis of CNN on multiple hardware platforms are instructive to provide references for practical engineering.In this thesis,a video pedestrian detection algorithm based on CNN is implemented on hardware,and the performance of algorithm implementation is analyzed.Our contributions are as follows:(1)Aimed at reducing the false negatives caused by motion blur and occlusion in videos,a postprocessing algorithm based on Kalman filter for SSD is proposed.Kalman filter propagates the detection results from SSD.It improves and stabilizes the confidence of pedestrian.The proposed algorithm balances the relationship among the miss rate,speed and the difficulty of hardware implementation.The proposed algorithm is tested to validate its effectiveness on both public and private datasets.Meanwhile,it also lays a foundation for the subsequent implementation on multiple hardware platforms.(2)Based on the analysis of CNN implementation requirements and the comparison of platforms,the hardware selection,CNN implementation and optimization on multiple hardware platforms are accomplished.Through testing and comparison,the implementation is validated as correct,it can detects pedestrian effectively.(3)The detection results of SSD and its post-processing algorithm on above platforms are analyzed from the two aspects of miss rate and speed.Based on the experiments and surveys,the limitations of each hardware platform are analyzed from the workload of network,power consumption and implementation difficulty,so as to provide references for engineering.In this thesis,the implementation and performance analysis of CNN on hardware platforms for video pedestrian detection are accomplished.The design of basic CNN components can be used in CNN implementation.The performance analysis provides the references for problems in engineering,such as algorithm design,hardware selection and algorithm implementation.