Research On Object Detection Based On Light-weight Convolutional Neural Networks In Traffic Scene

In recent years,autonomous driving has become a research hotspot worldwide,and ADAS(Advanced Driver Assistant System)in this field have been widely used.Object detection is a key component of this system.It is challenging to design an efficient object detection system in the scenario of autonomous driving where real time performance of object detection is critical.Because there are many small objects with varying amounts of occlusion occur.Object detection method based on deep convolutional neural network is adopted in this thesis and the public KITTI dataset is selected as the training set and test set.The main work of this thesis includes:(1)Aiming at the requirement in the real time performance and memory,an improved lightweight RetinaNet object detection algorithm is proposed.PeleeNet is used as a backbone network,then classification and regression subnets are optimized to improve detection speed and reduce the network complexity,so that the algorithm can be transplanted to hardware platform more conveniently.(2)Prior information about the specific dataset is used to design more reasonable anchors so that we can promote the detection accuracy further more.(3)The three-level features fusion and the attention mechanism is used,as a result,the overall detection performance of the model is further improved.(4)A detection model based on ASFF(Adaptive Spatial Feature Fusion)and PeleeNet-RetinaNet is proposed in this thesis,which reduced the amount of parameters and computation with higher detection performance.Experiments on the KITTI dataset show that the PeleeNet-RetinaNet detection model based on ASFF proposed improves detection performance and reduces resource consumption significantly when compared with RetinaNet based on ResNet-50.The number of parameters of the model is only 13.9% of RetinaNet based on ResNet-50,and the detection speed is 3.14 times faster.