Research On Lightweight Multi-Pedestrian Tracking Algorithm Based On DeepSort

Multiple Object Tracking aims to estimate the trajectory and motion state of object in video sequence.Due to extensive applicability among fields like autonomous driving and surveillance,multi-object tracking has been gathering more attention.This dissertation focuses on the research of Deep Sort(Simple Online and Realtime Tracking),a multiple object tracking algorithm following the tracking-by-detection paradigm,and aims to apply it to multi-pedestrian tracking task.Multiple objects tracking process consists of two parts under the paradigm of tracking-by-detection.Object detection is implemented first at the detection end,and then object trajectory will be predicted and associated with detected object on the strength of detection result at the tracking end.To accelerate tracking process without decrement of precision,improvements are proposed on both aspects of detection and tracking algorithms.The main content is summarized as follows:Firstly,as the fact that original Yolov3 has a relatively low precision,it is proposed to promote detection precision in terms of network structure and loss function.FPN(Feature Pyramid Network)is replaced with a more efficient structure.Coordinate attention mechanism and depth-wise separable convolution are introduced to further enhance detection precision.Loss function for regression is modified which helps speed up convergence of training and obtain a better regression precision of bounding box.Moreover,focal loss is introduced as classification loss to concentrate more on the learning of hard samples.With improvements mentioned above,a precision-enhanced model named v3-3 is obtained.To demonstrate the validity of proposed model,a new pedestrian dataset was collected by merging instances from different datasets and implementing data augmentation.Experiments show that the proposed model v3-3outperforms Yolov3 by 4.22% AP.Secondly,with an eye to the redundant computation of the proposed model v3-3,network pruning is implemented to ameliorate the depth-wise and width-wise redundancy.On the prerequisite of precision maintenance,a compact pruned model,namely v3-4,was obtained with appropriate channel pruning and layer pruning rate.Experiments on the pedestrian dataset and MOT16 dataset show that tracking algorithm with proposed pruned model,v3-4,achieves a tracking rate at 21.1 FPS,exceeding its counterpart Yolov3 by 11.4 FPS.Finally,as Kalman filter used in Deep Sort suffers from deficient robustness under nonlinear circumstance,in order to boost the robustness of Kalman filter,a multiple objects tracking algorithm based on adaptive Kalman filter is proposed.In the proposed method,standard Kalman filter is modified to adaptive Kalman filter which estimates noises according to innovations during tracking process in real time.It turns out that,on the MOT16 dataset,the proposed adaptive Kalman filter based algorithm outperforms its counterpart based on original Kalman filter by 2.05% and 1.13% in MOTA and MOTP,respectively.After changing different detectors,further experiments reveal that the proposed method bears validity in increasing the robustness of multiple objects tracking algorithm.