Online Forward Vehicle Detection Based On Multi-feature Fusion

As one of the key technologies of intelligent vehicles,the forward vehicle detection technology has been paid more and more attention,and how to improve the detection accuracy and detection speed has become a core problem.At present,detecting forward vehicle with single vehicle feature can no longer meet the accuracy,so in this thesis,multi-feature fusion method is used to deeply study the forward vehicle detection method.Firstly,the basic theory and extraction methods of single vehicle feature currently used are studied in this thesis,and the advantages and disadvantages of each single feature are analyzed.And based on the analysis results,Multi-channel Aggregated Features(MCAF)is proposed,and then,the maximum pooling is used to reduce the dimension of the features to solve the problem of large dimensions of the feature.According to the experimental results on the test dataset,using MCAF can obtain higher detection accuracy than single features and other fusion features.Secondly,forward vehicle detection model is build with LightGBM algorithm,and the key parameters affecting the performance of the algorithm are extracted according to the principle of this algorithm,and the optimization method combining one-by-one and grid search are used to optimize these parameters,so the best combination of parameters is obtained.Through the comparison experiments with other detection algorithms on the same test dataset,the superiority of this algorithm in training speed and classification accuracy is further verified.Finally,the online vehicle detection system based on quadratic classification and fast feature pyramid sliding window is adopted.The system is divided into two steps: the first step is to scan the entire image with small density windows and classify the windows with edge feature and LightGBM algorithm;The second step is to scan the region of interest extracted in step1 with the method of fast feature pyramid sliding window,then the windows are classified the second time with MCAF and LightGBM algorithm.The final detection result of forward vehicles is determined by combined all the positive windows which are classified with the algorithm.The feasibility and accuracy of the system framework is tested with road images and videos.In this thesis,the online forward vehicle detection model based on MCAF and LightGBM algorithm is build,and the comprehensive detection accuracy on the test dataset is 99.08%.In addition,the quadratic classification method is used to achieve 120 ms pre frame on detection speed,which basically meets the requirements of online detection.It provides a way for the forward vehicle detection technology,and has certain reference value.