The Design And Implementation Of Vehicle Detection System In Firefighting Scenario Based On Deep Learning

With the development of computer vision,object detection is one of the key tasks of computer vision.It is widely used in daily life,such as target tracking,autonomous vehicles,medical images analysis,military object detection,etc.Firefighting is an important part of national economic and social development.It is not only a guarantee for the development of a socialist market economy,but also related to the safety of people’s lives and property and social stability directly.This thesis combines the object detection technology in computer vision with the fire rescue mission.During the disaster,timely understanding of the traffic information around the fire as a reference for command and decision can not only help the fire rescue vehicle to arrive at the scene in the shortest time,but also provide assistance for the follow-up rescue work.This thesis collects traffic information through object detection techniques.Time is the most valuable rescue resource,so the algorithm that needs to be tested must satisfy the accuracy requirements and completes the detection of the object in real-time.This thesis introduces the development process of object detection technology from two directions: traditional object detection algorithm and the object detection algorithm based on deep learning.Combined with the actual application scenarios,then analyzes the commonly used object detection algorithms based on deep learning,finds out the problems that are not suitable for the scenario,and proposes solutions to different problems.For the problem of detection speed,this thesis compares various object detection algorithms,and finally chooses YOLOv3 which has good performance in speed and precision as the detection model.By referring to the YOLO-tiny network which has excellent performance in detection speed,the detection network structure of YOLOv3 is optimized to improve the detection speed.For the scale parameters of anchors that are not suitable for the dataset used in this thesis,it is analyzed in principle to obtain the appropriate anchor box parameters by using the K-means clustering method.Since many images in the scene are collected under poor lighting conditions,this thesis adopts the method of image preprocessing to ensure the accuracy of detection.Before the model is trained,the data from the real application scene is collected,annotated,and divided in to training set and test set.Finally,based on the improved object detection algorithm and the actual application scenarios,the thesis determines the requirements,and designs the overall architecture of the target detection system,then refines it into functional sub-modules and introduces the design of each module in detail.