| With the support of the new smart city construction and "one network for all" policy,various city services are being accelerated on the cell phone,and transportation,as the lifeline of a city,is the most important area of city services.With the economic growth in recent years,the number of cars owned by citizens has been increasing year by year,but it also leads to frequent traffic accidents.273,098 traffic accidents occurred in China in 2021,and the number of deaths was as high as 62,218,making traffic accidents the most common and deadliest accidental injury in China.Therefore how to rely on smart phones to reduce traffic accident casualty rate and improve driving safety has become the focus of this thesis.The most common cause of traffic accidents is driver fatigue,so driver fatigue detection is very important.However,even with fatigue detection,crashes may still occur.In this case,being able to quickly detect the occurrence of a crash and send a distress call to the outside world to help them get rescue and treatment in time can minimize casualties.However,there are some problems with the current mobile-based fatigue detection and crash emergency distress methods: the lightweight fatigue detection model leads to insufficient accuracy in order to ensure real-time detection;the traditional crash emergency distress takes the instantaneous location of the crash as the distress location,without taking into account the deviation of the final actual location of the victim being rescued from the instantaneous location of the crash.In response to the above two problems,this thesis carries out the following work:(1)For the problem of insufficient accuracy of fatigue detection models,this thesis uses the latest efficient lightweight neural network modules Mobile One and Mobile Former to replace its backbone network based on the PFLD model,respectively,and completes the network reconstruction through the re-parameterization and lightweight two-way cross-attention mechanism,and verifies the accuracy of these two models on the La Pa dataset by several metrics The accuracy of both models is verified by several metrics on the La Pa dataset.Finally,the PFLD-Mobile One model is selected under the consideration of performance and efficiency,which improves the performance by 16.8% and the inference speed by 30.4% compared with PFLD.(2)For the problem of emergency distress location deviation in car accidents,this thesis uses the Extended Kalman Filter(EKF)algorithm to fuse and process IMU and GNSS data,and optimizes the EKF parameters by the Particle Swarm Optimization(PSO)algorithm to track and predict the vehicle location in real time,so that the actual location of the victim can be sent out in time after a car accident.In the experiments,the control variables method is used to verify the accuracy of the algorithm on the EU Long-term dataset for high-precision sensors and on real roads for cell phone sensors for vehicle position tracking prediction.Finally,the smartphone equipped with the EKF-PSO algorithm predicted the position and direction of the vehicle after the collision more accurately using the RC model car to simulate a car crash.(3)In this thesis,the above two algorithms are implanted into the cell phone together with the voice interaction function through Android development technology.When the mobile phone detects that the driver has entered the fatigue state,it will remind the driver through the form of voice until the driver recovers awake.When the phone detects that the driver is in a car accident,it will actively ask the driver about the safety status through the voice assistant,and when the driver’s answer triggers the distress keyword,the phone voice assistant will make a response distress mechanism,such as notifying the police/hospital through SMS or phone or constantly beeping to try to wake up the victim’s consciousness.After testing,the average FPS of APP can reach 40.5,which meets the real-time requirements of fatigue warning and car accident emergency distress. |
PDF Full Text Request