Detection And Analysis Of Target Trajectory Based On Address-Events

Biological vision systems are much better than traditional digital systems in intelligence,resources,quality and power consumption.They use a completely different neuromorphic processing architecture and principle from traditional digital signal processors to achieve the purpose of identification,detection and tracking.Similar to biological system,the lower latency,higher dynamic range and lower power consumption give DVS(Dynamic Vision Sensor)great potential for vision sensing.Unlike traditional cameras that read the full sensor array to produce images at a fixed frame rate,event cameras can avoid processing of redundant information in traditional cameras,and solve the problem that traditional cameras hardly capture the motion trajectory of the high-speed moving object.Finally,we can deal with the research purpose of real-time trajectory detection and analysis with low power consumption and low latency,e.g.,tracking high-speed moving objects such as missiles and aircraft,autopilot and security monitoring,etc.Traditional image sensor records the total brightness values of pixel in the array during the exposure time,while DVS only reports change in pixel-level brightness over time.This is completely different from traditional image sensor that has the way of working referred to as “All pixel record data together”.As a result,DVS can filter redundant data from background and generate a detailed description for targets' motion trajectory,so as to solve the difficulties of the application for traditional object tracking algorithms in complex background.In other words,DVS can provide a complete path of targets motion no matter what the environment is.Under this kind of circumstance,this paper applies two algorithms based on address-events to detect and analyze the trajectory of moving targets.The first one is a novel address-events coherence detection algorithm.In this algorithm,through analyzing the information carried by the discrete spatial-temporal streams of events at a certain moment,the moving target is determined by generating multiple responses to the events simultaneously.Therefore,by filtering out irrelevant events and preserving coherent events,the algorithm calculates the coherence among events to obtain the position of moving target and implement detectionbased tracking.On the one hand,the innovations of this algorithm is to calculate directly on the spatial-temporal stream of events,and keep the original output signals of DVS while the concept of frame is abandoned.On the other hand,instead of iterating over each input event,the algorithm processes multiple events simultaneously,which ensures real-time object tracking.The principle of address-events coherence detection algorithm is simple and intrinsical,but it has relatively low controllability and universality.Taking advantage of the characteristics of the optical flow,the second algorithm realizes the object tracking based on optical flow driven by address-events.The main contributions of this algorithm are summarized as follows: first,the input signal is selectively activated during spike transmission period according to the characteristics of LIF neurons.Our algorithm introduces the concept of bionic LIF neurons on tracking speed in order to effectively filter redundant events and largely reduce the runtime of the algorithm.Furthermore,our algorithm successfully deals with the large memory consumption problem which most of algorithms nowadays exist.As a result,we can achieve the goal of the realtime tracking.Second,as for improving tracking accuracy,our algorithm respectively uses Gaussian weighting method on motion orientation and Low-pass weighting method on motion amplitude to correct optical flow vector.In order to overcome the shortcomings of existing algorithms which have low tracking accuracy and tend to loss the tracking target,the improved algorithm can achieve a more accurate object tracking.It not only reduces the probability of target loss,but also improves the robustness of object tracking.Third,the algorithm estimates the motion vector based on address-events during which any unstructured signal can be fitted by a local plane.This operation successfully avoids the problems of losing information “between frames” for a traditional camera,which will cause discontinuity of object tracking and make it unable to implement in any real situations.In addition,it makes the algorithm have stronger universality to variety of situations.Finally,in the object tracking field,the performance of the algorithm is verified by comparing with other state-of-art algorithms under multiple real scenarios,and the results show that we all get excellent performance.