Research On Deep Learning-based Multi-Object Detection In Complex Environment

Object detection based deep learning is a hot research direction in the field of computer vision.In recent years,excellent algorithms have been emerging,and Single Shot multi-box Detector,SSD is one of them.It has high detection accuracy and fast speed.It achieves high accuracy and real time detection of multi-object by using feature pyramids and the method of generating candidate regions of different scales and different aspect ratios on different layers.The resolution of the input images and the selection of the candidate region boxes are very important to the accuracy of the network.However,for a specific data set,if the original SSD model is used to select candidate region boxes,there will be redundancy and bias.In the current Single Shot multi-box Detector(SSD)algorithm,there are only two resolutions of input images,300×300 and 512×512.So in original SSD any input image will be adjusted to 300×300 or 512×512 size.However,the resolution of input images may be 1200×400 or higher.If these images are adjusted to 512×512,they will be distorted and the detection accuracy will be badly affected.In this paper,we propose a method to design a Personal SSD(P_SSD)which is suitable for specific resolution of input images to improve SSD algorithm to increase its classification accuracy.We modify the scales and ratios of default boxes by statistical learning method,and get the relevant parameters according to formula we designed.In addition,we replace the loss function in SSD with Focal Loss function to solve the problem of samples imbalance.Experiments on KITTI demonstrate that our method achieves competitive results in accuracy compared to SSD512,the mAP(mean average precision)is increased by 12%.A Driving Record Object Detection(DROD)dataset is also constructed by us to testify the effectiveness of our method.The experiment result based on DROD dataset shows that the mAP is increased by 10%compared to SSD512.The effectiveness of our improved SSD algorithm in this paper is verified once again.The improved SSD network can be used as a general object detection network model.It can be applied to any resolution data set,such as 2000 × 400,without worrying about the distortion of the image when it is input into the network.