Using Artificial Intelligence Chip And Deep Learning For Front-Vehicle Detection

Recently,with the increasing number of vehicles,traffic safety issues are increasingly prominent.Intelligent visual analysis technology can effectively detect possible safety issues and make corresponding control actions,thereby improving traffic safety issues.Front-Vehicle Detection is an important technology of intelligent visual analysis technology.The content we can detect includes various important information such as license plate number,car lights,seat belts,driver status,etc.,by detecting the head part of the car.It can issue alarms and feedback related information in time and take corresponding management measures when the set target violates the set safety rules.In actual application scenarios,the Front-Vehicle detection task requires real-time,accuracy,and low power consumption.At present,target detection has accuracy,but the speed is slow,far from real-time processing of camera data,and the detector consumes a lot of power when actually used.The artificial intelligence chip specially designed for the neural network structure can greatly speed up the calculation speed of the neural network,so that the front detection task meets the actual application and consumes lower power consumption.Therefore,this paper proposes the application research of deep learning methods for Front-Vehicle detection of artificial intelligence chips,including the following research contents:1.This thesis constructs a set of HD Front-Vehicle datasets under real scenes due to the fact that the current real traffic environment has less data.Data content covers:real-world weather environment,rich lighting changes;various vehicles;complex samples with multiple targets;complex image data such as reversing,straight ahead.A crossvalidation analysis was performed on the data set using the object detection algorithm based on regions under the Tensor Flow framework and the object detection algorithm based on border regression under the Py Torch framework.2.Considering the real-time requirements of the Front-Vehicle Detection task and the cross-validation results of the dataset,an end-to-end lightweight Front-Vehicledetector that combines multi-feature maps to extract features while using separable convolution to reduce the amount of parameters and calculations is designed;.At the same time,a visual analysis of the front detector was done.3.The end-to-end lightweight Front-Vehicle detector applied to the Cambrian MLU100,In the dataset constructed in this article,the speed is about 35 times higher than that of NVIDIA’s TITAN XP,while reducing power consumption.The performance of the Front-Vehicle detector under the MLU 100 single-card Cambricon caffe framework and the 32-core single-card multi-core mode with different data parallelism and model parallelism are discussed.4.The performance of a Front-Vehicle detector is comprehensively evaluated from three aspects of accuracy,speed and power consumption.The Cambrian MLU 100 made the above evaluations of the online mode and offline mode of the Front-Vehicle detector.