Research On Anti-collision Technology Based On Fusion Of Visual Perception And Millimeter Wave Radar

The anti-collision warning system can monitor the environment around the vehicle in real time and send early warning information to the driver in case of a possible collision with an obstacle,thus alerting the driver to avoid it,which can effectively reduce the occurrence of traffic accidents.Therefore,in recent years,the in-vehicle anti-collision system has become the research focus of domestic and foreign research institutions and major car companies.The sensors currently used in anti-collision systems mainly include cameras,millimeter-wave radars,and lidars.A single sensor has certain limitations.Cameras and lidars are susceptible to environmental factors such as climate.Although millimeter-wave radar can operate around the clock,it cannot perceive the type of target,and there is a certain chance of false alarms.To address this problem,this thesis designs a hardware system for the automotive aftermarket based on a vehicle-grade highperformance So C,and develops an anti-collision warning software system with multisensing fusion based on it.First,a deep learning SSD target detection network model was trained using the mature autonomous driving dataset KITTI,combined with samples from the selfcollection section,and the model was deployed on a vehicle-grade So C TDA2 S platform.The target detection algorithm was tested with a maximum frame rate of 25 fps and a detection accuracy of 96.5%.Second,targeting the vehicle aftermarket application requirements,this thesis designs a monocular vision all-in-one hardware system based on the TDA2 S chip,and designs supporting interface circuits and camera modules based on image and millimeter wave radar data acquisition requirements.Then,based on the model theory of the traditional small-aperture imaging camera,the internal and external parameters of the camera are calculated using the Zhang Zhengyou calibration method.Meanwhile,the rotation translation matrix of world coordinates and pixel coordinates is calculated based on the Pn P problem,which solves the measurement error problem of the traditional solution method and controls the coordinate transformation error within 20 pixels.Inspired by the multi-threaded programming semaphore,this thesis proposes a sensor time alignment algorithm to realize the time alignment of sensor sampling.The cross-parallel ratio（Io U）fusion algorithm is used to discriminate the target area after multi-sensor fusion to solve the image target false detection and millimeter wave radar false alarm problems.In order to improve the discrimination accuracy,the algorithm is optimized using millimeter-wave radar target reflection amplitude data.Since the system designed in this thesis is used in "dangerous and emergency" driving scenarios,the time to collision（TTC）is used as the basis for warning information generation,and the TTC threshold is set based on national standards and the experience of domestic and foreign ADAS manufacturers.Finally,in this thesis,the system is deployed in monocular vision all-in-one to test the performance of the anti-collision system with fusion of vision perception and millimeter wave radar.The system framework is designed based on Vision SDK,and the algorithm link of each module is designed and implemented according to the application requirements,while the system is tested statically and dynamically,and the accuracy rates of 94.6% and 90.7% are obtained respectively.