Lightweight CNN Model And Implementation Of FPGA Hardware Acceleration For Railway Pedestrian Detection

With the rapid development of the railway industry,driving safety has become one of the urgent problems to be solved.Among them,pedestrian intrusion is an important factor affecting driving safety.Therefore,timely detection of whether there are people who accidentally break into the track can ensure driving safety.However,traditional pedestrian detection methods are difficult to achieve high-performance and real-time detection tasks.In recent years,the object detection algorithm based on deep learning has achieved excellent performance.Therefore,this thesis is based on convolutional neural network,combined with embedded heterogeneous equipment to complete the pedestrian detection in the railway scene,and realize the intelligentization of railway security.The following are the main research work of this thesis.(1)Aiming at the problems of low detection accuracy and slow speed of traditional detection methods,a detection method based on deep learning is adopted to realize pedestrian detection.Based on the universal object detection algorithm YOLOv4,combined with the characteristics of railway scenes,a pedestrian detection algorithm dedicated to railway scenes is designed.In view of the difference between the scene of public dataset and the railway scene,a pedestrian dataset of local railway scenes was constructed on the basis of exiting database.Finally,the log-average miss rate,frames per second and average precision are used to evaluate the performance of the network.The results show that the algorithm has good detection results in the railway scenes.(2)Aiming at the problem of large pedestrian detection algorithm model,a lightweight pedestrian detection model is proposed.Based on the theory of channel pruning,this paper designs a strategy of pruning shortcut layer to ensure the regularity of the model after pruning.At the same time,the local attenuation sparsity training strategy is used to measure the importance of the network layer before pruning.Finally,the original dataset is used to fine-tune the pruned model.The experimental results show that the detection speed of the lightweight model is increased from 6.9 FPS to 9FPS under the premise that the logarithmic average missed detection rate is increased by less than 1%.(3)The deployment of lightweight pedestrian detection algorithms on hardware has been completed.The detection process of the three input methods of video stream,video file and image is realized.Multi-threaded scheduling is used to optimize video detection,and finally the performance of the pedestrian detection system is verified.This paper has completed the supplementary work of the dataset,the design of a lightweight network dedicated to pedestrian detection,and the deployment of the lightweight network in hardware devices.The results show that the deployed pedestrian detection system can better realize the detection function and can be used for pedestrian detection in railway scenes.