Research And Application Of Vehicle Recognition Based On Multi-Sensor Data Fusion

In major cities,the increasing of the car ownership has already brought a huge traffic problem.It is essential to develop intelligent transportation techniques to solve major challenges in the traffic problem.The vehicle recognition is an important part of the intelligent transportation system,and it is critical for intelligent traffic management to obtain vehicle type information on road timely and accurately.The conventional vehicle information detection has disadvantages of single individual method,incomplete information,which leads to low recognition accuracy and unstable recognition performance.To deal with these challenging problems,an integrated vehicle recognition method is proposed in this thesis combining coils,geomagnetism,and video sensors,which makes full use of advantages of each individual sensor.Also,this new method extracts features of the different sensors,and fuses them into multiple feature vectors.Using a mixed kernel with relevance vector machine to construct a vehicle classifier model,and it optimizes relevant parameters through particle swarm optimization.According to requirements of the following Jiangxi provincial science and technology project,“Design and Implementation of IoT Sensing Nodes for Intelligent Transportation Systems”managed by the research group,researches on the design of vehicle sensing nodes have been comprehensively conducted.Firstly,this thesis discusses the overall design of the vehicle perception system from the macro perspective.Then it describes the hardware design and software design of the sensor node.The OK6410 development board equipped with an embedded system of Linux 3.01 is used as a multi-sensor node hardware platform.To ensure the space-time synchronization of multi-sensor data,the time alignment and spatial alignment of sensing nodes are analyzed and designed.Finally,vehicle data has been collected and 11 types of coils,geomagnetism,and video sensors have been extracted through building up a sensory node in the field.Through the RVM model classifier introduced in this thesis,92.73% of vehicle model recognition accuracy has been achieved,and the proposed method has been compared with others based on their experimental results.Experimental outcomes demonstrate that,the relevance vector machine based on a mixed kernel,has better performance on this vehicle recognition and classification task.Compared with a single sensor,the multi-sensor data fusion vehicle recognition method has higher accuracy;and so,it has a good promotion and application value.