Research And Implementation Of Object Detectionand Tracking Algorithm Based On Deep Learning

Object detection and tracking has always been an important research direction in the field of computer vision,especially the problem of visual multi-target tracking has increasingly become the focus of computational vision research.With the development of deep learning and artificial intelligence,deep learning-based target detection and tracking has made breakthrough progress.In practical applications,due to cost and power consumption,the object detection and tracking algorithm is of great significance in the implementation of embedded platforms.Based on the SSD detection algorithm,this paper uses the DBT method to design vehicle and pedestrian multi-target tracking algorithms respectively.The vehicle tracking algorithm is deployed on the NVIDIA Jetson TX2 platform and the pedestrian tracking algorithm is deployed on the TI AM5749 platform.In terms of SSD detection,in order to improve the speed to adapt to the calculation of the embedded platform,this article uses a lightweight network to modify the basic network of the SSD to improve the detection speed.The training data set is filtered according to the actual application scenario.According to the distribution of vehicles and pedestrians in target samples,features are extracted on feature layers of different depths.Manually set the size,number and aspect ratio of the candidate boxes of each layer.On this basis,frame regression and classification prediction are performed to improve the detection accuracy.Finally,the average accuracy of vehicle detection trained on the BDD dataset reached 89.3%,and the average accuracy of pedestrian detection reached 61.6%.In terms of vehicle tracking,based on SSD vehicle detection,Kalman filtering is used to estimate vehicle motion status.Using residual network to extract different vehicle features.Combine the vehicle's motion information and appearance information to design the total associated cost.The Hungarian algorithm is used to complete the data association between the detections and the tracking trajectories,so as to realize the multi-target tracking of the vehicle.The complete detection-tracking system is deployed on the NVIDIA Jetson TX2 platform.In terms of pedestrian tracking,based on SSD pedestrian detection,combined with KCF tracking algorithm,a pedestrian tracking method is designed.SSD detection creates new tracking objects and corrects existing tracking objects.The location of the pedestrian target is given from the tracking result to complete the pedestrian multi-target tracking.The detection-tracking algorithm is implemented on the TI AM5749 platform.In terms of implementation,the caffe-jacinto framework is used for L2 regularization training,L1 regularization training,and sparse training.The network weight parameter sparsity reached 60% and the weight parameters were quantized by 8 bits.The network model is grouped and the DSP core and EVE core are used for network inference calculation.The KCF tracking algorithm is deployed on the ARM core for operation.The development of detection and tracking applications is achieved through multi-core allocation.