Research On Pedestrian And Vehicle Detection Technology Based On Deep Learning

With the continuous increase of the number of traffic accidents in China and people’s attention to travel safety,how to reduce traffic accidents has aroused the attention of the automobile manufacturing industry.Driverless cars have developed rapidly in recent years and will become one of the ways to reduce traffic accidents in the future.Target detection based on image information is one of the key technologies in driverless vehicles.Target detection algorithm based on deep learning is widely used in environmental awareness tasks due to its high detection accuracy and strong robustness.However,the current target detection algorithm based on deep learning still has many areas to be optimized.Pedestrian detection and vehicle detection,due to the small target,the pixel of the target changes greatly,there are missed detection and wrong detection problems in detection.In this paper,according to the general target detection algorithm itself in the face of the defects of the application scene and pedestrian and vehicle detection problems,through the model lightweight and the introduction of better feature extraction network to solve the above problems.The main research contents of this paper include:Pedestrian detection and vehicle detection experiments are carried out on the common target detection algorithm.The advantages and disadvantages of the common target detection algorithm in pedestrian and vehicle detection tasks are studied by comparative analysis.Aiming at the problem that the number of parameters in SSD algorithm greatly affects the operation speed of the algorithm,a lightweight neural network feature extraction structure was designed based on dense connection and separable convolution.The comparative experiments show that the number of parameters in the optimized algorithm is greatly reduced,and it has better detection accuracy and detection speed in the pedestrian detection task.To solve the problem of missed detection in pedestrian detection due to the small size of the target,this paper proposes to improve the detection accuracy of the algorithm for small targets by integrating more shallow information,and to expand the receptive field and compress the depth of the network by means of void convolution.Based on BIFPN,a feature fusion structure with 3 inputs and 2 outputs is designed,and the detection accuracy is improved by superposition of feature fusion layer.Experimental results show that the optimized algorithm in this paper is better than Yolov3 algorithm in pedestrian detection task.Due to the large amount of post-processing in the target detection algorithm based on prior box,the algorithm complexity is too large.In this paper,a feature extraction network for the target detection algorithm without prior box is designed by introducing the channel attention mechanism and improving the Inception structure in the Resnet network.The network is applied to the Centernet no-prior box target detection algorithm to realize vehicle detection,and the experiment shows that the algorithm has a higher detection accuracy in vehicle detection task.