Vehicle Detection Based On Deep Learning

Thanks to the rise in popularity of deep learning,numerous breakthroughs are made in Computer Vision.Being one of the low-level tasks,object detection has a wide variety of downstream applications in the industry.As a result,industry is seeing increasing demands and attempts to deploy methods with improved performances.Among all these down-stream applications,vehicle detection is known as the one with great potential and experiences rapid development in the applications such as carparking detection,traffic violation detection,autopilot etc.Despite the promising progress on the model effectiveness(accuracy,precision)in the academic field,the question is raised whether a method can transfer well into the industry field.As we know,the efficiency(calculation speed and resource cost)of a framework weighs equally,if not more heavily,when it comes to model application and deployment.This paper modifies the existing strategies from the perspectives of both effectiveness and speed,using the cars images from COCO dataset.In summary,the main contribution of this work are as follows:(1)I introduce network architecture search(NAS)method orienting detection.Conventional NAS methods design and optimize the network architecture on basic image classification tasks.However,due to discrepancies between tasks,the networks optimized for these basic image classification tasks show fewer satisfying performances in car detection.To address this issue,this paper directly searches for backbone network based on car detection task.Using a super network establishing action search space,differentiablize the latency and add it to the searching metrics.In the objective function,I set both precision and speed as constraint terms at the same time.Eventually a network architecture of unchanged precision is obtained,with an increased speed of 16%.This method is capable of searching on different hardware platforms’ characteristics,in the meantime it takes delay rate into consideration when predicting network architecture.The search process does not involve additional hardware.This hardware-in-circle feature means that the calculation cost is saved by a remarkable amount.(2)I introduce an anchored,one-stage and multi-scale vehicle detection algorithm.To address the sizing problem in vehicle detection,I introduce a feature-leaning fused,multi-scale,self-adaptive learning module to ensure model’s capability to capture the multi-scale features.What is more,I search the most suitable focal loss by grid for the task of vehicle detection,and it is called Resample Loss.(3)I introduce an anchor-agnostic vehicle detection algorithm.To compress the network architecture further,I just modify the head network structure to remove the anchor box module and instead adopts the detection technique based on key point regression.Thus I alleviate the balancing issue of positive and negative samples induced by region proposal being too dense.I add one more centerness branch based on regression subnet,which can decrease the number of low-quality boxes without any extra cost.