Research On Obstacle Detection Of Outdoor Mobile Robot Based On Deep Learning

When the outdoor unmanned vehicle drives autonomously,it needs to detect obstacles in the environment to decide the next action.However,for the research of outdoor obstacle detection,the traditional image recognition algorithm still have the problems of slow detection speed and low detection accuracy.In recent years,the rapid development of deep learning ideas in the field of computer has solved these problems.In this paper,the experiment platform is based on an unmanned vehicle,by monocular vision sensors to record of unmanned vehicle outdoor environment,and based on the deep learning algorithm,the outdoor obstacle detection system is designed to solve the problem that the existing algorithm has low accuracy in detecting small target obstacles.The main research contents include the following aspects:(1)In SSD algorithm,after the image passes through the front network layer,the loss of feature information is large and the location prediction is not accurate.It is improved by changing the pooling layer of VGG-16 network and adding H-type expansion convolution.With the improvement of the image and the improved resolution of the network,the resolution of the feature graph is smaller,the precision network output,and the accuracy of the algorithm to detect the target.(2)For the problem that the network structure of SSD algorithm is not sensitive enough to small-sized targets,in order to improve the detection accuracy of small target obstacles,the network structure was improved after VGG-16.The improved network can make full use of the characteristics of each network layer by using the improved method of input mapping and input superposition to enter the lower convolution.(3)For the problem that the original SSD algorithm is easy to cause training interruption when training the network,the improved algorithm uses SeLU as the activation function of the network.SeLU can achieve self-normalization,prevent gradient explosion or gradient disappearance in the training process,and guarantee the robustness of neural network learning.This paper started from the actual application situation,took three common outdoor obstacles:car,person,bicycle and self-made data sets as training samples,and used data augmented principles to augment the data.The experiment used SGD Momentum algorithm and the method of vector loop for model trainingTested on the test set,the average detection accuracy(mAP)of inputting 300 X 300 size pictures was 80.4%,which was 6.3 percentage points higher than the original algorithm,and the speed reached 33 frames/second.The paper evaluated the retrieval ability of small targets,and designed several control tests to analyze the improved model.The retrieval rate for ultra-small targets and small targets is 0.58 and 0.87,which is 20 and 13 percentage points higher than the original algorithm.The experimental results showed that the improved algorithm effectively improved the detection accuracy of the system for small target obstacles.