Research And Improvement Of Object Detection Algorithm Based On Regression Method

With the development of convolutional neural networks,as well as advances in hardware technology and the arrival of age of big data,the object detection technology not only greatly improves the detection rate,but also greatly accelerates the detection speed.This allows object detection to be implemented in systems that require real-time detection.Such as vehicle-oriented system for object detection tasks through the camera to capture the road vehicle of image information,automatic accurate real time detection of images of cars,pedestrians and cyclists,namely recognition and localization.The traditional object detection algorithm needs to extract features manually,which leads to the problem of poor robustness and low average detection rate,and it is difficult to detect multi class objects.The object detection algorithm based on sliding window or region proposal needs post-processing,and there is a problem that the detection speed can not reach the real time and the memory demand is so big.The object detection algorithm based on the regression method has many advantages,such as fast detection speed and low memory requirement,which can detect multi class objects.However,the detection rate of object detection algorithm based on regression method is still low,especially for small objects and obscured objects.In order to solve these problems,this paper improves the object detection algorithm based on regression method,and improves detection rate and recall rate on the premise of ensuring real-time speed.The SqueezeDet algorithm has the fastest detection speed on the KITTI data set,and is also one of the object detection algorithms which have the highest detection rate based on regression method.This paper makes improvements based on the SqueezeDet algorithm,then improve the average detection rate.In the structure of the network,the shortcut connections structure with gate function,not only abstract expression in high feature space,but also enhance the detection of small objects and obscured objects using the details in shallow feature space;in the loss function,the width and height of the bounding box will be normalized to a relative change instead of the variance,and then improve the imbalance problem of bounding box loss optimization for large objects and small objects,and in order to improve the classification accuracy,will also add the incorrectly classified with penalty to classification loss calculation;in the bounding box filtering,using the non-maximum suppression method with weighted to solve the redundancy problem of the bounding box in the baseline.In addition,the network in baseline has a little layers and much pooling operations,because of the pooling operation causes the details hardly to retain,small objects and obscured objects with low detection rate,so using dilated convolution operation replace the pooling operation,improve the receptive field without changing the resolution.The experimental results show that the average detection rate by our method on the KITTI data set is 80.1%,the recall rate is 81.6%,and the detection speed is 0.032 s.Compared with the baseline,the average detection rate is increased by 3.4%,the recall rate is raised by 2.8%,and the detection speed is raised by 2.7 ms.This method is also porting programs to mobile devices to verify the applicability and real-time performance of the algorithm.