Research On Hybrid Vehicle Target Detection Technology Based On Deep Learning

Recently,the method of detecting mixed pedestrian and vehicles in roads has become a key point in the construction of intelligent transportation systems and has received extensive attention from domestic and foreign researchers.The target detection tasks involved in intelligent transportation mainly include capturing the position information of pedestrians,vehicles,and other hybrid targets on the currently monitored road segment at a certain time and frequency by the use of road camera equipment.In the actual road scenario,the background is complex,the attitudes of pedestrians and vehicles are different,and the distribution range of the target scale is extremely wide.Therefore,there are many problems in the detection of hybrid targets in this scenario.At the same time,the existing classical target detection algorithms are mostly based on deep learning,but such methods generally have higher computational complexity.Therefore,it is a great challenge to study efficient and reliable hybrid target detection algorithms in road scenarios.In order to solve the above problems,this paper designs and implements a hybrid target detection method for intelligent traffic system of people and vehicles that can perform hybrid multi-target detection for real road scenarios,and has improved accuracy and speed compared to previous detection algorithms.The specific work is as follows:1)According to the scale distribution of the target in the real road scenario,the scale ratio of the original candidate frame is optimized.In this way,when the candidate boxes are non-linearly operated through the multi-layer network,it is easier to return to the target itself.2)To further solve the problem of multi-scale distribution of mixed targets,this paper uses multiple receptive field areas with different sizes to adaptively match the different scales of the target,and concatenates the convolutional layers from the network structure so that the size of the scale is not identified.One of multiple pedestrian vehicles mixes goals.3)For the problem of high computational complexity in the hybrid target detection algorithm,this paper uses several small convolution kernels with strong non-linear expression capability to replace part of the full connected layer.As a result,the network connection parameters can be greatly reduced without losing the non-linear expression capability of the network,which greatly improves the speed and the consumption of the memory resources.This paper takes the shared convolution features as the starting point to study the adaptability of the receptive field and candidate frame scales for multi-scale target detection,and then solves the problem of detecting mixed targets under real road scenarios.Using the above optimization method,we achieved good results in the road scenario.After a long-term test on the site of a high-speed road in a certain city,Device detection accuracy increased by 7.79%,detection speed increased by 10FPS,and test memory reduced by 2245M,which fully proves that the target detection method for specific road sections designed in this paper can meet the accuracy and stability of the actual application.