Research And Implementation Of Object Detection Based On Road Monitoring Video In Complex Environment

Object detection has always been a very popular research direction of computer vision,and has a wide application prospect.At present,traditional object detection algorithms have encountered bottlenecks and deep learning algorithm is also applied to object detection of single image.However,most algorithms pay more attention to the accuracy,rather than the real-time.At the same time,there are more and more application scenarios for target detection,and the demand for target detection in complex environments is increasing.The goal of this thesis is to achieve real-time object detection in complex scenarios with accuracy.First,we improves the feature extraction in the classical object detection algorithm,and we improves the algorithm to meet the real-time detection of the project while ensuring accuracy.Then the new model is trained on the deep learning algorithm and a new post-processing algorithm is proposed to complete the multi-object detection task.The traditional machine learning algorithm selected in this thesis is classic SVM algorithm.HOG of edge information are input into the SVM for training.A large amount of data is not required for training and the training time is very short.Training has no special requirements of hardware,but SVM has problems such as poor generalization ability and low classification accuracy.Deep learning algorithms far exceed traditional machine learning algorithms in terms of accuracy and generalization ability.Object detection algorithms based on deep learning can achieve object detection in complex environments.However,deep learning algorithms have disadvantages of needing a large amount of data for training,the training process is long,and the hardware equipment require of high training requires.This thesis completes the object detection based on SVM and CNN algorithm respectively.Work and innovations of this thesis are as follows:1)We implement real-time pedestrian and vehicle detection based on SVM.A new edge extraction algorithm is proposed to process the input image.New regularization template HOG method is used to extract the edge information.We use the the data to train the SVM to detect object.The experimental results show that the real-time performance of the improved algorithm and strategy detection is significantly improved,and the accuracy is slightly improved.2)The algorithm based on deep learning completes the task of multi-target detection.In this thesis,considering the balance between real-time and precision,SSD model is selected.We improve the detection accuracy by adding pre-training of complex scene data,retraining the fully connected layer,and proposing a new post-processing algorithm.The trained model in this thesis is applied to intelligent road monitoring while 11 common classes object on the road are detected.Experimental results show that the improved model of this thesis has higher detection accuracy,higher detection rate,and can cope with complex scenes such as rain,haze,and dark light.