Research On Vehicle Detection Application Based On Ascend NPU And Lightweight Convolutional Neural Network

The accuracy of the general target detection convolutional neural network algorithm continues to improve,and with it comes the explosive growth of the amount of calculation and the amount of parameters.In contrast,the emerging fields of intelligent transportation,autonomous driving,and intelligent security all use embedded AI devices,which poses new challenges for convolutional neural network algorithms.In-depth compression of highperformance target detection networks can greatly reduce the amount of network parameters.This is essential to deploy target detection convolutional neural networks on embedded AI devices that lack computing resources.The Ascend NPU uses a customized Da Vinci architecture.How to realize the fast convolution calculation process and give full play to the advantages of the Ascend NPU is the key to vehicle detection applications.On the basis of common target detection,a lightweight feature extraction network is used as the backbone network to design and train a lightweight vehicle detection network with high detection accuracy and small parameters.Aiming at the problem of uneven feature distribution before and after low-precision quantization,Kullback-Leibler divergence is used to measure the difference in information distribution before and after quantization,and the quantization mapping interval is calibrated.Aiming at the problem of wide distribution of scale parameters in the batch normalization layer during the pruning process,the L1 regularization is used to constrain the scale parameters,and the network is sparsely trained to improve the efficiency of pruning.Aiming at the byte alignment problem of data transmission in Ascend NPU,a five-dimensional data transmission format is designed,and the convolution data is tiled to realize fast convolution calculation.For the ordinary convolution calculation process,the use of multi-stage pipeline and multi-core parallel technology to further increase the calculation speed.The experimental results show that the vehicle detection of the lightweight backbone network achieves a compression rate of 93% on Center Net and only a loss of accuracy of4.89%.The 8-bit quantization based on KL divergence calibration reduces the accuracy of the YOLO v4 network model by only 0.87%,and the model size is reduced by four times.The sparse pruning method further compresses the model by 65% under the lightweight Center Net model.The final vehicle detection model is only 2.64 MB,and the detection accuracy after pruning is almost unchanged.After Center Net accelerated by multi-stage pipeline and multi-core in parallel,the inference time on the Ascend NPU reaches 4.547 ms per frame,which far exceeds the speed requirement of real-time detection.