Research And Application Of Real-Time Recognition Based On Video Compressed Domain

Surveillance video-based object recognition technology has gradually become popular in all aspects of both industry and daily life.With the increasing capabilities of surveillance equipment,the quality of surveillance videos,such as resolution and frame rate,has also been improving.The above status quo directly leads to a sharp increase in the amount of surveillance video data volume,prompting a continuous increase in the computational complexity of video-based analysis.This puts tremendous pressure on the limited hardware resources and eventually leads to a significant increase in computational latency,making it virtually impossible to perform tasks in real-time.Take the field of intelligent transportation as an example.Its demand for surveillance video analysis,such as abnormal vehicle identification,and driving speed estimation,is increasing.In the meanwhile,the real-time guarantee of the analysis algorithms is vital to road safety.To solve the delay problem incurred by traditional pixel-domain video analysis algorithms,this thesis takes the measurement of vehicle speed as the application background,and focuses on the compressed-domain-based real-time object recognition algorithm and its implementation in embedded devices.The main research contents of this thesis are as follows.(1)A compression domain-based information extraction method is proposed to address data volume and processing delay issues in conventional pixel domain-based object detection and recognition approaches.The proposed method completes target marking by processing the video compressed domain information and analyzing only this information,thus avoiding the full decoding of the surveillance video.The proposed method first partially decodes the video stream to obtain an initial motion vector field,then performs temporal-spatial denoising on the field to obtain a reliable motion vector field,and finally binarizes the denoised field to generate a target mask,before utilizing the mask to locate the target.By comparing this method with representative pixel-domain neural-network-based methods,i.e.,YOLO and FASTER R-CNN,it is shown that the processing speed of the proposed method is substantially better than that of the neural network methods with an average processing speed of only about 14 ms while satisfying the minimum accuracy requirement.(2)To address the problems that the current pixel domain-based target tracking methods are computationally intensive and susceptible to tracking errors due to environmental factors such as light and motion blur,a motion vector-based tracking approach is proposed,which takes the target frame obtained from the previous research content based on the compressed domain information as the input,predicts the target’s movement over a period through the motion vector,before tracking is performed based on the prediction result.After tracking is completed,the pixel-domain velocity of the tracked object is further calculated by first using feature point analysis on the tracked object to extract the feature points of the target,and the pixel displacement of the target between adjacent frames is calculated by the pixel distance between multiple successfully matched feature points,combined with the video frame rate.For the matching error,the tailed mean removal method is used to reduce its impact on the final result.Finally,the pixel displacement of the pixel-domain coordinate system is transformed into the physical world,completing the real-world speed estimation of the tracked target.(3)For the vehicle speed measurement application scenario,this thesis deploys the proposed method on embedded devices with limited computational power.Through further data compression and optimization,an embedded system for speed measurement is realized.The system demonstrates vehicle speed measurement capabilities with acceptable accuracy requirements.The device is small in size,easy to install,and highly compatible,and can be combined with other task systems to accomplish various tasks related to real-time target identification and speed measurement.In summary,this thesis designs a compressed-domain-based object recognition method,which replaces the traditional pixel-domain data analysis by analyzing the compressed-domain data of the recognized targets,reduces the resource consumption of decoding,decreases the amount of data computation,and greatly reduces the analysis latency under acceptable accuracy requirement.The study finally realizes a low-latency vehicle speed measurement system based on road surveillance video,which can still accomplish the speed measurement task despite the resource constraint.The system provides a new video-based speed measurement solution for intelligent transportation,which exhibits substantially lower delay than traditional pixel domainbased speed measurement methods,and substantially lower maintenance costs than the dedicated-hardware-based approaches.