Research On Video Target Detection Algorithm Based On Light-weight Neural Network

Target detection is an important part of automatic driving system.The object detection method based on machine learning has achieved remarkable results.However,in the autonomous driving environment,high storage resources and computing resources cannot be provided.The traditional video target detection algorithm is not suitable for automatic driving environment due to its complex network structure.Therefore,it is necessary to design a video target detection algorithm with high accuracy and real-time performance.The video target detection algorithm is improved as follows:To solve the problem of low computing resources in automatic driving environment,a light detection network was proposed.Since the feature extraction network requires a large amount of computation,the attention-based channel pruning is proposed.The normalized layer is used to reduce the channel and optimize the loss function.The attentional mechanism model based on residual network is designed to add global information.Since the regression network parameter redundancy is high,a fine-grained pruning based on weight sharing matrix is proposed.Optimize the weight sharing matrix and select the appropriate regularization.The optimal pruning rate and fine-tuning times were determined by experiments,and the number of parameters was reduced by 33%.Significantly reduce the number of network parameters.Aiming at the problem of high real-time requirement of video target detection algorithm in automatic driving environment,a video preprocessing algorithm is designed.The video is divided into key frame and non-key frame,and the detection network only detects the target of the key frame.Reduce the time consumption of video frame detection.After the video is divided into key frame and non-key frame,the occlusion fuzzy target is missed.A feature propagation network based on non-key frame features is proposed,which integrates the features of key frame and non-key frame to enhance the feature extraction of key frame.The efficient detection of occlusion,blur and special attitude in video target detection is realized.Through cascade detector and optimized detection post-processing algorithm.Increase the training samples of small targets and enhance the detection of small targets.All optimization stages are verified by experiments based on the new open source data set of the four-dimensional diagram.The proposed video target detection algorithm ensures the accuracy of the original network and improves the detection speed.