Research And Implementation Of Vehicle Pedestrian Detection System Based On Jetson TK1

In recent years,with the gradual rise of ADAS(advanced driving assistance system)and autonomous driving technology,as the key technology autonomous driving,pedestrian detection has become a hot issue in the field of computer vision.Pedestrian detection system can automatically identify pedestrians in dangerous areas at longer distances position and alarm drivers to prevent the occurrence of the accident,so it can greatly reduce the probability of traffic accidents.However,due to the complexity of the environment where the pedestrians are usually in a variety of circumstances,and the diversity of the vehicle and the autonomous vehicle has a large volume and high power consumption.So far there is not a more general and mature approach to solve problems such as low accuracy,poor real-time performance and excessive power consumption,so in the practical application of the vehicle pedestrian detection system opportunities and challenges coexist.In this paper,Jetson TK1 embedded device with low power consumption and high performance parallel operation is proposed as the hardware platform,and the Jetson TK1 based on the deep learning convolution neural network algorithm is proposed to solve the difficulties in the above-mentioned.Firstly,the paper introduces the overall design framework of the system,and introduces the hardware acceleration platform and the deep learning framework of caffe based on Jetson TK1.The hardware part mainly introduces the CPU and GPU acceleration based on Jetson TK1 and the CUDA programming mode GPU acceleration principle;software part of the main introduction to caffe and its advantages.Secondly,according to the existing pedestrian detection algorithm,the traditional pedestrian detection algorithm based on manual feature training classifier of HOG + SVM is introduced.Due to the shortcomings of high false positive rate and poor real-time performance,this paper proposes a pedestrian detection algorithm based on convolution neural network.In this paper,the CNN algorithm is used to optimize the network by using the whole convolution layer instead of the whole connection layer on the original classical model,so that the network structure can input any size picture,and avoid the repeated storage and convolution calculation using the pixel block problem,making the model more efficient.At the same time,K-means algorithm is used to cluster the data sets,and the number of convolution cores of the output layer is further reduced,which reduces the model parameters and improves the calculation speed of the model.In order to verify the accuracy and real-time performance of this algorithm,HOG + SVM algorithm and the CNN algorithm are used to train and test the same INRIA pedestrian data set.Theexperimental results show that the algorithm has a high true positive rate and has a faster detection speed in the low power embedded device,which can meet the practical requirements of the application.