Research Of Real-time Vehicle And Traffic Sign Detection On Android Platform Based On Tensor Flow

In the process of driving a car,due to the driver's negligence,many traffic signs are often undiscovered,and because of the driver's inattention,the vehicles and pedestrians in front of the vehicle can not be found to cause rear-end accidents.In traffic driving,the road environment is complex.A driving assistance system that can detect important traffic information and alert drivers can effectively reduce the occurrence of violations and accidents,and provides a safe driving solution.Real-time detection of vehicles,pedestrians and traffic signs is actually a real-time object detection task.With the rapid improvement of the performance of computer hardware equipment,and the continuous development of deep learning and object detection technologies,high-rate object detection can be achieved with high accuracy and real-time object detection results.In recent years,the object detection technology based on deep learning has developed rapidly,and different object detection models have been continuously proposed.Some of the latest object detection technologies can ensure both high detection accuracy and high detection speed.Moreover,the current high performance of mobile devices can meet computing requirements of the deep learning,and making it possible to run target detection on the Android platform.This research develops and improves an object detection model based on deep learning to achieve real-time detection of vehicles,pedestrians and traffic signs on the Android mobile platform which is based on the TensorFlow machine learning platform.In this research,based on the TensorFlow platform,the SSD model is improved and created based on SSD(Single Shot MultiBox Detector)and combined with other technologies.Sample images are collected to create training data sets which are used to train models,and the dynamic-link library of the TensorFlow model is compiled to be invoked on the mobile platform.Design and develop an Android program,and in the Android program the camera is used to continuously capture the preview image in the front of the vehicle,call the trained model to detect the image,and display and broadcast the detection result.In the experimental project,according to the characteristics of the target task,use different schemes to improve the object detection model to improve the accuracy and speed of detection.In the research,the following improvements are made to the SSD model: using the MobileNetV2 deep neural network as the basic network of the SSD model;using the K-Means clustering algorithm to set the default box in the SSD;using the feature rearrangement under sampling and deconvolution,the channel fusion produces a new feature map structure.The test result of the experiment shows that: The improvement of the SSD model improves the accuracy of the detection,and the design and implementation of the system achieves the desired results.The program running on the Android platform can detect objects such as vehicles and traffic signs in real time.The results of the project are practical in assisted driving and are referenced to other object detection technologies.