Research And System Implementation Of Vehicle Object Detection Method Based On Improved SSD Algorithm

Vehicle object detection has drawn a lot of interest in real-world traffic situations because of its capacity to swiftly and correctly identify moving objects.This has made it easier to manage traffic flow,ease congestion,and boost traffic efficiency.Most conventional algorithms for vehicle detection can only handle single-road scenarios,making it challenging to manage complicated contexts like target occlusion,target truncation with changing light conditions,and other road variables.Also,the identification is not as accurate and quick as needed for practical application due to the weak resilience of the manually designed features.This topic conducts research on vehicle target identification algorithms based on deep learning techniques in order to obtain effective and intelligent vehicle detection,as follows.(1)In order to address the issue of inadequate identification of small and truncated targets in real traffic settings,an enhanced SSD method is proposed in this study.With an attention mechanism added to the feature extraction phase to better the extraction of shallow feature information,Resnet50 serves as the improved SSD algorithm’s backbone network.Recursive Feature Pyramid(RFP)iterative feature pyramid structures are used to increase the fusion of features between multi-scale feature maps.The experiments demonstrate that RFP-SSD outperforms the original SSD algorithm in terms of detection efficiency using the KITTI dataset for training.In comparison to the SSD algorithm,the RFP-SSD algorithm gets mAP metric of 87.77%,6.77% improvement.(2)This study suggests a mobile device-friendly lightweight car detection algorithm.The initial SSD backbone network is reconfigured using the fused network created by this algorithm,which also merges the ECA module and MobilenetV2.The computational effort of the feature extraction network is considerably reduced when the reconfigured backbone network is used to extract features from vehicle images.Then,an iterative feature pyramid structure is introduced to fuse the deep features of the light-weighted backbone network to improve the efficiency of the model’s fusion of deep features.The experimental results show that the model has an average detection accuracy of 80%,a detection speed of 52 frames and a model size of 32.6MB,which is a better performance compared to the comparison model.(3)This paper designs and develops a vehicle detection system with picture detection and video detection capabilities.The system includes both vehicle image detection and video detection functions,and it can display the detection results immediately after receiving an image or video input,improving the visibility of the algorithmic detection effect.In conclusion,the light-weighting of the improved RFP-SSD algorithm using the MobilenetV2 network results in a reduction in model complexity and an increase in detection speed,while also meeting the demands for real-time detection on mobile hardware devices,lowering detection costs,and having good practical value.