Design And Implementation Of Vehicle Detection Algorithm For Assisted Driving

With the development of technology,vehicles have become the main means of transportation for people.However,while the vehicles give great convenience,rapidity and comfort to people's travel,traffic accidents become more and more frequent due to the driver's carelessness and so on.So,vehicle safety has gradually became a focus of people's concern.The forward collision warning system with high performance can effectively detect vehicles and remind drivers whether the vehicles in front and around will collide with the current driving vehicle,which has the realistic meaning with importance for reducing the possibility of traffic accidents.As the core part of forward collision warning system,vehicle detection algorithm has made great progress due to the rapid development of convolution neural networks(CNN).However,it is difficult to be used in the embedded system for the problem that there are too many parameters in the CNN.And vehicle detection algorithm based on CNN has massive computations,so it can't achieve the real-time requirements of vehicle assisted driving products.Therefore,this paper studies the vehicle detection algorithms based on CNN and proposes a vehicle detection network model which can achieve real-time detection on the embedded platform.The research lays the foundation for the application of vehicle detection for vehicle assisted driving.The main research work of this paper are as follows:1.We propose an improved vehicle detection algorithm based on SSD.Firstly,the inception block is introduced into the network structure,so that the network can obtain more image information.Secondly,we redesign the scales of the default bounding boxes.So,the scales of default bounding boxes in shallow feature maps have smaller intervals,while the intervals in deep feature maps become larger.Furthermore,the K-Means algorithm is employed to obtain the clustering values of the aspect ratios of the vehicle samples.The clustering value is used as the initial value of the aspect ratios of the default bounding boxes,which reduces the number of default bounding boxes.Experiments were carried out on KITTI and Union Vehicle Detection(UVD)datasets,and the results show that compared with SSD,the method proposed in this paper can not only improve the average detection accuracy of vehicles,but also the design of default bounding boxes is more suitable for vehicle position regression.2.To further compress the improved vehicle detection network,the parallel method of filter pruning and parameter quantization is adopted.The method of ThiNet is used to prune filters.According to the information guidance of the i+1th layer,the filter of the layer i is pruned.ThiNet uses a subset of the input of the i+lth layer to approximate the output of the i+1th layer and trims out other unrelated filters.Parameter quantization is to use K-Means algorithm to get the effective weights of each layer.Multiple connections share the same effective weight to reduce the number of weights which are need to be stored.The experimental results show that the parallel method of filter pruning and parameter quantization can achieve a compression rate of 15.9%with little effect on the average detection rate,and the speed of network detection has been improved nearly third.3.We transplant the improved network on the TX2 platform to test all aspects of the network,which can further study the application of the compressed network on the embedded platform.The improved and compressed network is transplanted to the embedded platform in 17MB,and we test the detection accuracy and speed of the network under the various actual driving environments.The experimental results show that the network can reach the speed of 19 FPS on the embedded platform with a high average detection accuracy.So,it can meet the functional and performance requirements of vehicle assisted driving products.