Research And Implementation Of Pedestrian Flow Statistics Embedded System Based On Machine Vision

With the development of computer vision,intelligent monitoring has developed rapidly.Recently,increasing demand to process data locally is put forward with the improvement of hardware performance and emergence of edge computing.To deal with low accuracy and poor real-time performance of machine vision algorithm on embedded platform,in this paper technology related to pedestrian traffic statistics is studiedFirstly,an embedded system for pedestrian traffic statistics based on machine vision is designed according to system requirements and application scenarios with the algorithm flow of "pedestrian detection-pedestrian tracking-pedestrian counting",and its software and hardware platform is built.Then,the pedestrian detection algorithm of pedestrian traffic statistics embedded system is studied.The SSD Deep Learning Detection Framework is selected with an improved default box with the consideration of special pedestrian target.At the same time,a Top-Down structure is proposed to enrich the spatial and semantic information of feature maps at different levels and then improve the detection accuracy.An improved backbone network based on deepthwise separable convolution is proposed to improve the real-time performance under the embedded platform with limited computing resources.The results on Caltech test set show that the improved algorithm decreases all and reasonable pedestrian miss rates by 12.14% and 4.92%,respectively with the detection speed at 19.7FPS,and The improved algorithm balances the performance of detection accuracy and speed compared with other mainstream detection algorithms,suitable for embedded platform applications.Secondly,the pedestrian tracking algorithm of pedestrian traffic statistics embedded system is studied.The KCF tracking algorithm based on correlation filtering is proposed with multi-feature,FHOG,CN and gray feature,fusion strategy to improve discriminability,with a scale filter to adapt the target scale change,and with a template update strategy based on multi-peak detection to improve the tracking robustness of occluded targets.The results on OTB dataset show that the overall accuracy and success rate of the proposed algorithm are 0.752 and 0.623,respectively with the tracking speed at 41.3 FPS.So the improved algorithm is applicable to the application scenario and embedded platform for its tracking accuracy and practicability.Finally,the embedded system of pedestrian traffic statistics is tested and analyzed.The detection results are correlated with the tracking results to delete the targets beyond the visual field and add the new targets in the visual field.Then,combined with trajectory analysis method,pedestrian targets count by the analysis on the target.The Result in actual scenarios shows that the average recognition rate is 93.30% and the average frame rate is 18.25 FPS.The pedestrian traffic statistics system proposed in this paper has strong practicability improving accuracy and real-time in embedded platform.