Non-motor Vehicle Detection Based On Deep Learning

With the development of artificial intelligence technology and Internet of Things technology,unmanned driving and vehicle-road coordination technologies have been developed by leaps and bounds.At the same time,sustainable development is receiving more and more attention in social progress.Green travel modes such as public transportation,walking and cycling are encouraged,and the safety of non-motor vehicles on the road needs to be guaranteed.The non-motor vehicle detection system based on deep learning can detect non-motor vehicles in the motor vehicle flow on the road,and share the detection results with the motor vehicles through the vehicle-road coordination system to improve road safety.The current road monitoring system is mainly aimed at motor vehicles,and the amount of data of non-motor vehicles is small and the targets of non-motor vehicles are also small.In response to the above challenges,the data preprocessing technology adopts geometric transformation methods such as mirror transformation,stretching and compression transformation,color transformation methods such as brightness,chroma,contrast transformation,and superimposed rain and snow motion trajectories for data enhancement to simulate various light and weather conditions for surveillance situations and expand the data set.In view of the low density of non-motor vehicle targets,the target detection model based on the sparse design concept combines mainstream and cutting-edge target detection methods,abandoning the heavy Region Proposal Network,and only provides a small number of candidate frames for each input image,and using the feature fusion method to enhance the feature expression.The model removes the repeated Non-Maximum Suppression process,and adopts the iterative detection method,it means that the positioning and classification results by the last layer can be input to the next layer to achieve refinement,so that the detection network can be accurate.The data is collected from the real road environment of the road section near the school,and the experiment is completed based on the preprocessed data set.In order to obtain the most suitable model for non-motor vehicle detection tasks in the motor vehicle flow,an experiment was designed based on changed iterative detection structure and changed initial number of candidate frames.The experimental results show that the model can always converge well,and has a good performance on target detection tasks.Among them,under a reasonable number of training times,the model with 6 iterations of detection head and the initial embedding of 100 candidate frames performs best.When the Io U threshold is set to 0.5,the accuracy can reach 91.8.No matter the detection times of single image equals 10 or 100,the recall rate can always reach more than 87%,and the processing speed of single image is less than 0.12 seconds.