Research On Target Detection Tracking And Collision Avoidance Warning In Blind Area Of Heavy Vehicles

The large size of heavy vehicles leads to the existence of visual blind area around the vehicle body,which is an important reason for human-vehicle collision casualty accidents.The existing solutions to improve the driving safety of heavy vehicles include installing blind spot mirrors and wide-angle mirrors to expand the vehicle’s blind area viewing angle,installing crash barriers around the body of heavy vehicles,delineating blind area warning strips in specific areas of the road,and following the instructions of people outside the vehicle.However,there are relatively few applications and studies on safe driving assistance methods for heavy vehicles.Based on the existing theories and methods,aiming at some key research problems such as the accuracy and real-time performance of dynamic target detection in the right blind area of heavy vehicles,the precision of dynamic target tracking,the ranging accuracy and stability of static targets in the rear blind area of heavy vehicles,and the feasibility of blind area anti-collision early warning methods,this paper carries out the research on target detection tracking and collision avoidance warning in blind area of heavy vehicles.It is of great significance to reduce the number of heavy vehicle traffic accidents and reduce the casualty rate.The main research work of this paper is as follows:1.Constructing sample data based on optimized generative adversarial network data augmentation method.Aiming at the lack of cyclist sample data,a sample data augmentation method based on optimized generative adversarial network is proposed,which is used for the first time to generate cyclist sample images.This paper optimizes the deep convolutional generative adversarial network in terms of loss function and network structure,further generates more abundant sample data based on a small amount of real sample data of cyclists and pedestrians,and constructs the real sample image training set and the augmented data training set.The experimental results show that the stability of the optimized generative adversarial network is further improved,and the quality of the generated images is better,which provides a feasible solution for the augmentation of training sample data.2.A lightweight convolutional network dynamic target detection method based on gradient sparse factor is proposed,and a heavy vehicle right blind area dynamic target detection dataset is established.The impact of augmented data on network detection performance is evaluated under the YOLOv5 framework,and the location loss function and post-processing method of YOLOv5 are improved.A sparse training method based on gradient sparse factor is proposed,which performs sparse pruning and lightweight processing on the network model.Establish a heavy vehicle right blind area dynamic target detection data set to evaluate the experimental effect of dynamic cyclist and pedestrian detection in the actual environment.The experimental results show that data augmentation and network improvement can effectively improve the accuracy of model detection,lightweight processing can effectively reduce the model size and parameters,and improve model detection speed.3.By fusing the shallow and deep features of dynamic targets,a dynamic target tracking method in the right blind area of heavy vehicles based on multi-feature representation and correlation filtering is proposed.The motion trajectory of the dynamic target in the right blind area of the heavy vehicle is studied,combined with traditional features of FHOG,LBP,CN and deep features extracted by Res Net-50 for multi-feature description.Multi-feature fusion is performed through interpolation processing,and dynamic target tracking is performed using correlation filter algorithm.Expand the dynamic target tracking data set in the right blind area of heavy vehicles to test the performance of dynamic target tracking methods in the actual environment.The experimental results show that the multi-feature representation can effectively improve the precision and AUC of the tracking method,and the improvement effect of deep features is better than that of shallow features.4.Based on distance data fusion of multi-sensor and federated unscented particle filter,a research on static target ranging in the rear blind area of heavy vehicle is carried out.A multienvironmental perception system based on binocular stereo cameras,ultrasonic radars and laser rangefinders is constructed to obtain ranging information of static targets in the rear blind area of heavy vehicle,and the federal unscented particle filter algorithm framework is used for distance data fusion.The experimental data is collected in the actual environment of the blind area behind the heavy vehicle,and the simulation results show that the method can effectively improve the accuracy and stability of the static target ranging in the blind area behind the heavy vehicle.5.Heavy vehicle blind area collision avoidance warning model and strategy are proposed.Based on the Bayesian network collision avoidance early warning model for heavy vehicles and the reversing collision avoidance early warning strategy based on fuzzy control theory,this paper conducts a simulation analysis of heavy vehicle blind area collision avoidance early warning to improve the driving safety of heavy vehicles.Simulation experiments verify the feasibility of the heavy vehicle blind area collision avoidance warning model and strategy.This paper studies the safety of the right and rear blind area of heavy vehicles,analyzes some key research problems such as heavy vehicle blind area target detection tracking and collision warning.According to the existing theoretical and technical research,a targeted solution is given,which playes a positive role in improving the driving safety performance of heavy vehicles.